Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
AI ndash Sep2018 ndash Alisson Sol ndash [1]
AI circa 2018
bull How to solve quadratic equationbull 1198861199092 + 119887119909 + 119888 = 0
bull Who is learning better in new airplane mission simulator
bull What is the next number in the sequence11 21 1211 111221
Pictures Wikimedia Commons license or taken by author
Pilot Week 1 Week 2
Olivia 0 75
Oliver 25 100
bull While we wait to start think about problems in the righthellip
AI ndash Sep2018 ndash Alisson Sol ndash [2]
Quadratic equation solution
bull From Algebra
bull 119909 =minus119887plusmn 1198872minus4119886119888
2119886
bull Why do we care
AI ndash Sep2018 ndash Alisson Sol ndash [3]
Going to higher ordershellip
bull Cubic equationhellip Anyonebull 1198861199093 + 1198871199092 + 119888119909 + 119889 = 0
bull Specific examplehellipbull 1199093 minus 21199092 minus 5119909 + 6 = 0
bull What ifhellipbull We plot it
bull Or could 119904119900119897119907119890 119894119905
AI ndash Sep2018 ndash Alisson Sol ndash [5]
Evolution of AI
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [6]
Which AI
bull Todaybull Applied AI (aka ldquoweak AIrdquo or ldquonarrow AIrdquo)
bull Out of scopebull Artificial General Intelligence aka ldquostrong AIrdquo or ldquofull AIrdquo
bull Consciousnessbull Like in movies and TV series 2001 A Space Odyssey AI Her Humans Westworld hellip
AI ndash Sep2018 ndash Alisson Sol ndash [7]
Numerical Calculus Newton (~1685-1740)
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [2]
Quadratic equation solution
bull From Algebra
bull 119909 =minus119887plusmn 1198872minus4119886119888
2119886
bull Why do we care
AI ndash Sep2018 ndash Alisson Sol ndash [3]
Going to higher ordershellip
bull Cubic equationhellip Anyonebull 1198861199093 + 1198871199092 + 119888119909 + 119889 = 0
bull Specific examplehellipbull 1199093 minus 21199092 minus 5119909 + 6 = 0
bull What ifhellipbull We plot it
bull Or could 119904119900119897119907119890 119894119905
AI ndash Sep2018 ndash Alisson Sol ndash [5]
Evolution of AI
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [6]
Which AI
bull Todaybull Applied AI (aka ldquoweak AIrdquo or ldquonarrow AIrdquo)
bull Out of scopebull Artificial General Intelligence aka ldquostrong AIrdquo or ldquofull AIrdquo
bull Consciousnessbull Like in movies and TV series 2001 A Space Odyssey AI Her Humans Westworld hellip
AI ndash Sep2018 ndash Alisson Sol ndash [7]
Numerical Calculus Newton (~1685-1740)
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [3]
Going to higher ordershellip
bull Cubic equationhellip Anyonebull 1198861199093 + 1198871199092 + 119888119909 + 119889 = 0
bull Specific examplehellipbull 1199093 minus 21199092 minus 5119909 + 6 = 0
bull What ifhellipbull We plot it
bull Or could 119904119900119897119907119890 119894119905
AI ndash Sep2018 ndash Alisson Sol ndash [5]
Evolution of AI
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [6]
Which AI
bull Todaybull Applied AI (aka ldquoweak AIrdquo or ldquonarrow AIrdquo)
bull Out of scopebull Artificial General Intelligence aka ldquostrong AIrdquo or ldquofull AIrdquo
bull Consciousnessbull Like in movies and TV series 2001 A Space Odyssey AI Her Humans Westworld hellip
AI ndash Sep2018 ndash Alisson Sol ndash [7]
Numerical Calculus Newton (~1685-1740)
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [5]
Evolution of AI
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [6]
Which AI
bull Todaybull Applied AI (aka ldquoweak AIrdquo or ldquonarrow AIrdquo)
bull Out of scopebull Artificial General Intelligence aka ldquostrong AIrdquo or ldquofull AIrdquo
bull Consciousnessbull Like in movies and TV series 2001 A Space Odyssey AI Her Humans Westworld hellip
AI ndash Sep2018 ndash Alisson Sol ndash [7]
Numerical Calculus Newton (~1685-1740)
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [6]
Which AI
bull Todaybull Applied AI (aka ldquoweak AIrdquo or ldquonarrow AIrdquo)
bull Out of scopebull Artificial General Intelligence aka ldquostrong AIrdquo or ldquofull AIrdquo
bull Consciousnessbull Like in movies and TV series 2001 A Space Odyssey AI Her Humans Westworld hellip
AI ndash Sep2018 ndash Alisson Sol ndash [7]
Numerical Calculus Newton (~1685-1740)
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [7]
Numerical Calculus Newton (~1685-1740)
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [8]
1 second
2315 days
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [9]
Usual development
Output
Input
Processor
Code
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [10]
Computational power enabled brute forcehellip
bull k-Nearest Neighborsbull Pros accuracy insensitive to outliers little ldquodata preparationrdquo
bull Cons computationally expensive
bull Basic idea ldquoTell me who you walk withhelliprdquo
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [11]
Movies data
Source book Machine Learning in Action by Peter Harrington
Movie of kicks of kisses Type
A 3 104 Romance
B 2 100 Romance
C 1 81 Romance
D 101 10 Action
E 99 5 Action
F 98 2 Action
18 90 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [12]
Movies Scatter chart
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
119889 = (119909119886 minus 119909119887)2+(119910119886 minus 119910119887)2
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [13]
Movies Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
A 3 104 Romance 205
B 2 100 Romance 189
C 1 81 Romance 192
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [14]
Movies Sorted Distance
Movie of kicks of kisses Type d
18 90 Unknown 00
B 2 100 Romance 189
C 1 81 Romance 192
A 3 104 Romance 205
D 101 10 Action 1153
E 99 5 Action 1174
F 98 2 Action 1189
if kNN = 3NN thenhellip
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [15]
What If (BI Scenarios)
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [16]
Pre-calculated ldquoborderrdquo
A
R
R
R
R
R
R
R
R
R
R
R
R
A
A
R
R
R
R
R
R
R
R
R
R
R
A
A
A
R
R
R
R
R
R
R
R
R
R
A
A
A
A
R
R
R
R
R
R
R
R
R
A
A
A
A
A
R
R
R
R
R
R
R
R
A
A
A
A
A
A
R
R
R
R
R
R
R
A
A
A
A
A
A
A
R
R
R
R
R
R
A
A
A
A
A
A
A
A
R
R
R
R
R
A
A
A
A
A
A
A
A
A
R
R
R
R
A
A
A
A
A
A
A
A
A
A
R
R
R
A
A
A
A
A
A
A
A
A
A
A
R
R
A
A
A
A
A
A
A
A
A
A
A
A
R
A
A
A
A
A
A
A
A
A
A
A
A
A
0
20
40
60
80
100
120
0 20 40 60 80 100 120
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [17]
Cluster representation
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
CentroidsRomance = [295]Action = [9933567]
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [18]
Right method and correctly implemented
A Few Useful Things to KnowAbout Machine Learningby Pedro Domingos
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [19]
Training is costly
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [20]
Data analysis for an experiment
From Kinect for Windowspresentation duringMicrosoft Build 2013 event
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [21]
What is this
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [22]
What is Deep Learning
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [23]
What is deep learning
bull Deep learning (also known as deep structured learning hierarchical learning or deep machine learning) is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation
bull But what is a Neural Network
bull Visualize httpplaygroundtensorfloworg
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [24]
Who is learning better
bull Two pilots have been learning how to complete a difficult mission in an airplane simulator Rates of success per week are in the table
Pilot Week 1 Week 2 Aggregated
Olivia 0 75
Oliver 25 100
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [25]
Simpsons paradox
bull Trend in different groups of data disappears or reverses when these groups are combined (aka reversal or amalgamation paradox)
Pilot Week 1 Week 2 Aggregated
Olivia (01) 0 (34) 75 (35) 60
Oliver (14) 25 (11) 100 (25) 40
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [26]
Accuracy paradox
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Accuracy
bull119860 =119879119875+119879119873
119879119875+119865119875+119879119873+119865119873
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [27]
Model comparison by accuracy
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119860(119860) =100 + 9700
100 + 150 + 9700 + 50= 098 119860(119861) =
1 + 9849
1 + 1 + 9849 + 149= 0985
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [28]
F1 Score
Confusion matrix
PredictivePositive
PredictiveNegative
Positivesamples
True Positive
False Negative
Negativesamples
False Positive
True Negative
Measurements (Precision Recall F1)
bull119875 =119879119875
119879119875+119865119875
bull119877 =119879119875
119879119875+119865119873
bull1198651 = 2119875lowast119877
119875+119877
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [29]
Model comparison by F1 Score
Model A
PredictivePositive
PredictiveNegative
Positivesamples
100 50
Negativesamples
150 9700
Model B
PredictivePositive
PredictiveNegative
Positivesamples
1 149
Negativesamples
1 9849
119875 = 04 119877 = 067 1198651 = 05 119875 = 05 119877 = 00067 1198651 = 0013
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [30]
Takeaways Math and AI
bull Computers evolved enabling expanded use of Numerical Calculus
bull Someone in the team needs to understand the Math
bull Math may be right yet its interpretation may be incorrect
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [31]
Any questions before we go to ldquoMemoryrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [32]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [33]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [34]
这是什么
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [35]
What just happened
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [36]
Your mindhellip
bull Analyzed images
bull Split it into components
bull Associated such components with a verbal description
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [37]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [38]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [39]
你有这段记忆吗
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [40]
Playing with Image Search
bull httpsimagesgooglecom
bull httpswwwbingcomimages
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [41]
Your thoughts
bull How important are the features for the search
bull ldquoAn image is worth a thousand wordsrdquobull Would any 1000 words be equivalent
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [42]
Movies another dataset
Movie Running Time (Minutes) Revenue ($ million) Type
A 88 407 Romance
B 129 1811 Romance
C 119 4634 Romance
D 116 129 Action
E 102 534 Action
F 108 332 Action
115 15 Unknown
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [43]
Movies Scatter chart kicks and kisses
AB
C
DEF
0
20
40
60
80
100
120
0 20 40 60 80 100 120
o
f ki
sse
s
of kicks
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [44]
Movies Scatter chart revenue and duration
A
B
C
DE
F
0
100
200
300
400
500
80 90 100 110 120 130 140
Rev
en
ue
($
mill
ion
)
Running Time (Minutes)
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [45]
Orphaned MLAI projects
bull Pattern for failed ML projectbull Data was accumulated Volume Variety Velocity Veracity
bull Model was built with ldquopotentialrdquo success for initial anecdotes
bull Then different questions or need for more precise answers
bull Result project collapses
bull Corollary data is thrown away or lostbull httpstoolboxgooglecomdatasetsearch
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [46]
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [47]
Personal data collection half-marathons
00000
03000
10000
13000
20000
23000
30000
0 2 4 6 8 10 12 14
Seattle2015
Portland2015
SF2014
LakeSam2016
SFRnR2016
CDA2016
SEARnR2016
Portland2016
LARnR2016
LakeWA2016
AZRnR2017
LakeSam2017
Mercer2017
SEARnR2017
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [48]
Takeaways Memory and AI
bull Human memory is still poorly understood
bull Try the reverse of current trendbull Start with the questions
bull Build a model (proof of concept)
bull Then seek for the data
bull Iteratehellipbull Everything questions model data
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [49]
Any questions before we go to ldquoModelsrdquo
bull Math
bull Memory
bull Models
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [50]
Humans and modeling our intuition is bad
bull Intelligence ldquohellip can be more generally described as the ability to perceive or infer information and to retain it as knowledge to be applied towards adaptive behaviors within an environment or contextrdquo
bull Earth = center of universe flat
bull Birds fly have feathers hellip from Icarus to Leonardo da Vinci
bull Humans donrsquot understand the ldquofeaturesrdquobull Sports stock market relationships hellip
bull Models include abstractions ldquogravityrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [51]
Cognitive biases
httpsenwikipediaorgwikiList_of_cognitive_biases
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [52]
Biases and modeling
bull Decision-making belief behaviorbull DunningndashKruger effect
bull Irrational escalation (sunk cost)
bull Parkinsons law of triviality (bikeshedding)
bull Social biasesbull Illusory superiority (Lake Wobegon effect)
bull Fundamental attribution error
bull Memory errors and biasesbull Bizarreness effect
bull Hindsight bias
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [53]
How is Physics applied
bull Acceptance of multiple modelsbull No wrong or right model the best predictor wins the day
bull Acceptance of imprecise answersbull Simplifies models eliminating a lot of noise
bull 119865 = 119898119886 and 119864 = 1198981198882 only for ideal conditions
bull Detaching models from databull Nobel Prize in Physics 2017 Rainer Weiss Barry C Barish Kip S Thorne
LIGO Laser Interferometer Gravitational-Wave Observatoryconfirmed predictions by Albert Einstein 100 years ago
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [54]
Applying AI
bull Do you need to solve the general problem
bull Going from applied ML to applied AI
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [55]
Airplane autopilot
bull Shiprsquos gyroscopic-compass set 1242065 Elmer A Sperry
bull Automatic pilot for airplanes1707690A Lawrence B Sperrybull Control of position course or
altitude of land water air or space vehicles eg automatic pilot with safety arrangements for transition from automatic pilot to manual pilot and vice versa
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [56]
Applied Machine Learning
bull Get a goal
bull Make a goal-related model
bull Collect data
bull Test hypothesis
bull Apply prediction
bull Maximize flight occupation
bull Model for no-shows
bull Gather ticket info
bull Check predictions for no-shows
bull Actions based on predictionsbull Reminderspenalty for no-show
bull Allow more overbooking
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [57]
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [58]
End-to-end trip workflow
1 Luggage + documentation
2 Passenger gets to airport
3 Security checkpoints
4 Boarding
5 Flight + service
6 Leaving plane + connections
7 Luggage + customs + immigration
8 Airport to destination
9 Unpackhellip
1 Ouibring + documentation
2 Uber
3 Clear TSA PreCheck
4 Waiting room + priority check-in
5 Classes of service
6 Class of service + connections
7 Priority luggage + entry systems
8 Uber
9 Unpackhellip
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [59]
Workflows and ldquointelligencerdquo
bull Intelligence will come from optimizing the end-to-end workflow
bull You wanted to play musicbull Before buy tapevinylCD put media on device locate song press play
bull Then the media disappeared song went directly to your iPod
bull Then full process was streamlined ldquoAlexa play [song]rdquo
bull Conversational user interfacebull Trying on assistant Google httpsassistantgooglecomexplorehl=en_us
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [60]
What if airlines couldhellip
bull Predict flights by a class of passengers (holidays Super Bowl)
bull Offer full package flight accommodation etc
bull Alert that a passport is about to expire ahead of trip
bull Send transportation to customer houses
bull hellip
bull Summary anticipate and optimize complete job end-to-end
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [61]
Models and data
bull Which modelsdata can be sharedbull Models about passenger behavior
bull Models about airline operations (Crew equipment supply chain)
bull Models about services
bull Should models be ldquoregulatedrdquobull EU PNR (Passenger Name Record)
bull What about auditing
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [62]
AI and Ethics
bull Why is this a topic at all
bull Security and privacy
bull Will the intelligence change the worldmodel
bull How could intelligence maximize revenue for a hospital
bull Overbooking scenario which passenger would have to wait
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [63]
Takeaways Models and AI
bull Requirementsbull Machine scalability
bull Organizational expertise
bull Data
bull Experiments
bull Libraries (software)
bull ldquoThink big and long termrdquo
hellip perceiveor inferhellip
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221
AI ndash Sep2018 ndash Alisson Sol ndash [64]
QampA
bull Math
bull Memory
bull Models
bull Alisson SolemailAlissonSolcom
bull What was the next number in the sequence11 21 1211 111221