Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
DNNConnect2016
Artificial Intelligence with DNN
Jean-Sylvain Boige
Aricie
jsboigeariciefr
DNNConnect2016
Please support our valuable sponsors
DNNConnect2016
Summary
bull Introduction to AI
bull What is AI
bull Agent systems
bull DNN environment
bull A Tour of AI in DNN
bull Problem Solvingbull Search optimization
bull Reasoningbull Logic Knowledge bases
bull Dealing with uncertaintybull Probabilistic networks
bull Machine learningbull For all of the above
DNNConnect2016
Motivation
bull Aricie- Portal Keeper
bull Agents in DNN
bull httpsariciepkpcodeplexcom
bull Future My Intelligence Agency
bull You(rs)
bull Course at ORT engineer school
bull httpwwwredditcomria101
bull Artificial IntelligenceA Modern Approach
DNNConnect2016
What is AICS 4055
Turing Computing Machinery and Intelligence
Views of AI fall into four categories
Most useful approach acting rationallyrdquo
Build agents
Thinking humanly Thinking rationally
Acting humanly Acting rationally
DNNConnect2016
Computer
Science amp
Engineering
AI
Mathematics
Cognitive
Science
Philosophy
Psychology Linguistics
BiologyEconomics
Foundations of AI
bull 1943 McCulloch amp Pitts Boolean circuit model of brain
bull 1950 Turings Computing Machinery and Intelligenceldquo
bull 1956 Dartmouth meeting Artificial Intelligence adopted
bull 1950s Early AI programs Logic Theorist Geometry Engine
bull 1965 logical reasoning
bull 1970s AI discovers computational complexity
Neural network research almost disappears
bull 1970s Early development of knowledge-based systems
bull 1980s AI an industry Robotics Planning Control theory
bull 1986 Neural networks return to popularity
bull 1990s AI becomes a science
bull 1995 The emergence of intelligent agents GAs Artificial Life
bull 2000s Bayesian learning Knowledge Engineering
bull 2010s Deep learning ndash Smart contracts
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Please support our valuable sponsors
DNNConnect2016
Summary
bull Introduction to AI
bull What is AI
bull Agent systems
bull DNN environment
bull A Tour of AI in DNN
bull Problem Solvingbull Search optimization
bull Reasoningbull Logic Knowledge bases
bull Dealing with uncertaintybull Probabilistic networks
bull Machine learningbull For all of the above
DNNConnect2016
Motivation
bull Aricie- Portal Keeper
bull Agents in DNN
bull httpsariciepkpcodeplexcom
bull Future My Intelligence Agency
bull You(rs)
bull Course at ORT engineer school
bull httpwwwredditcomria101
bull Artificial IntelligenceA Modern Approach
DNNConnect2016
What is AICS 4055
Turing Computing Machinery and Intelligence
Views of AI fall into four categories
Most useful approach acting rationallyrdquo
Build agents
Thinking humanly Thinking rationally
Acting humanly Acting rationally
DNNConnect2016
Computer
Science amp
Engineering
AI
Mathematics
Cognitive
Science
Philosophy
Psychology Linguistics
BiologyEconomics
Foundations of AI
bull 1943 McCulloch amp Pitts Boolean circuit model of brain
bull 1950 Turings Computing Machinery and Intelligenceldquo
bull 1956 Dartmouth meeting Artificial Intelligence adopted
bull 1950s Early AI programs Logic Theorist Geometry Engine
bull 1965 logical reasoning
bull 1970s AI discovers computational complexity
Neural network research almost disappears
bull 1970s Early development of knowledge-based systems
bull 1980s AI an industry Robotics Planning Control theory
bull 1986 Neural networks return to popularity
bull 1990s AI becomes a science
bull 1995 The emergence of intelligent agents GAs Artificial Life
bull 2000s Bayesian learning Knowledge Engineering
bull 2010s Deep learning ndash Smart contracts
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Summary
bull Introduction to AI
bull What is AI
bull Agent systems
bull DNN environment
bull A Tour of AI in DNN
bull Problem Solvingbull Search optimization
bull Reasoningbull Logic Knowledge bases
bull Dealing with uncertaintybull Probabilistic networks
bull Machine learningbull For all of the above
DNNConnect2016
Motivation
bull Aricie- Portal Keeper
bull Agents in DNN
bull httpsariciepkpcodeplexcom
bull Future My Intelligence Agency
bull You(rs)
bull Course at ORT engineer school
bull httpwwwredditcomria101
bull Artificial IntelligenceA Modern Approach
DNNConnect2016
What is AICS 4055
Turing Computing Machinery and Intelligence
Views of AI fall into four categories
Most useful approach acting rationallyrdquo
Build agents
Thinking humanly Thinking rationally
Acting humanly Acting rationally
DNNConnect2016
Computer
Science amp
Engineering
AI
Mathematics
Cognitive
Science
Philosophy
Psychology Linguistics
BiologyEconomics
Foundations of AI
bull 1943 McCulloch amp Pitts Boolean circuit model of brain
bull 1950 Turings Computing Machinery and Intelligenceldquo
bull 1956 Dartmouth meeting Artificial Intelligence adopted
bull 1950s Early AI programs Logic Theorist Geometry Engine
bull 1965 logical reasoning
bull 1970s AI discovers computational complexity
Neural network research almost disappears
bull 1970s Early development of knowledge-based systems
bull 1980s AI an industry Robotics Planning Control theory
bull 1986 Neural networks return to popularity
bull 1990s AI becomes a science
bull 1995 The emergence of intelligent agents GAs Artificial Life
bull 2000s Bayesian learning Knowledge Engineering
bull 2010s Deep learning ndash Smart contracts
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Motivation
bull Aricie- Portal Keeper
bull Agents in DNN
bull httpsariciepkpcodeplexcom
bull Future My Intelligence Agency
bull You(rs)
bull Course at ORT engineer school
bull httpwwwredditcomria101
bull Artificial IntelligenceA Modern Approach
DNNConnect2016
What is AICS 4055
Turing Computing Machinery and Intelligence
Views of AI fall into four categories
Most useful approach acting rationallyrdquo
Build agents
Thinking humanly Thinking rationally
Acting humanly Acting rationally
DNNConnect2016
Computer
Science amp
Engineering
AI
Mathematics
Cognitive
Science
Philosophy
Psychology Linguistics
BiologyEconomics
Foundations of AI
bull 1943 McCulloch amp Pitts Boolean circuit model of brain
bull 1950 Turings Computing Machinery and Intelligenceldquo
bull 1956 Dartmouth meeting Artificial Intelligence adopted
bull 1950s Early AI programs Logic Theorist Geometry Engine
bull 1965 logical reasoning
bull 1970s AI discovers computational complexity
Neural network research almost disappears
bull 1970s Early development of knowledge-based systems
bull 1980s AI an industry Robotics Planning Control theory
bull 1986 Neural networks return to popularity
bull 1990s AI becomes a science
bull 1995 The emergence of intelligent agents GAs Artificial Life
bull 2000s Bayesian learning Knowledge Engineering
bull 2010s Deep learning ndash Smart contracts
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
What is AICS 4055
Turing Computing Machinery and Intelligence
Views of AI fall into four categories
Most useful approach acting rationallyrdquo
Build agents
Thinking humanly Thinking rationally
Acting humanly Acting rationally
DNNConnect2016
Computer
Science amp
Engineering
AI
Mathematics
Cognitive
Science
Philosophy
Psychology Linguistics
BiologyEconomics
Foundations of AI
bull 1943 McCulloch amp Pitts Boolean circuit model of brain
bull 1950 Turings Computing Machinery and Intelligenceldquo
bull 1956 Dartmouth meeting Artificial Intelligence adopted
bull 1950s Early AI programs Logic Theorist Geometry Engine
bull 1965 logical reasoning
bull 1970s AI discovers computational complexity
Neural network research almost disappears
bull 1970s Early development of knowledge-based systems
bull 1980s AI an industry Robotics Planning Control theory
bull 1986 Neural networks return to popularity
bull 1990s AI becomes a science
bull 1995 The emergence of intelligent agents GAs Artificial Life
bull 2000s Bayesian learning Knowledge Engineering
bull 2010s Deep learning ndash Smart contracts
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Computer
Science amp
Engineering
AI
Mathematics
Cognitive
Science
Philosophy
Psychology Linguistics
BiologyEconomics
Foundations of AI
bull 1943 McCulloch amp Pitts Boolean circuit model of brain
bull 1950 Turings Computing Machinery and Intelligenceldquo
bull 1956 Dartmouth meeting Artificial Intelligence adopted
bull 1950s Early AI programs Logic Theorist Geometry Engine
bull 1965 logical reasoning
bull 1970s AI discovers computational complexity
Neural network research almost disappears
bull 1970s Early development of knowledge-based systems
bull 1980s AI an industry Robotics Planning Control theory
bull 1986 Neural networks return to popularity
bull 1990s AI becomes a science
bull 1995 The emergence of intelligent agents GAs Artificial Life
bull 2000s Bayesian learning Knowledge Engineering
bull 2010s Deep learning ndash Smart contracts
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
AI in your everyday life
bull Post Office
bull address recognition bull sorting of mail
bull Banks
bull check readersbull signature verificationbull loan application
bull Customer Service
bull voice recognitionbull Planning
bull The Web
bull Identifying your age gender location from your Web surfing
bull fraud detection
bull Digital Cameras
bull face detection and focusing
bull Computer Games
bull Intelligent charactersagents
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Agents and environments
Agent function maps from percepts history into actions
[f P A]
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Simple reflex agents
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Model-based reflex agents
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
DNN and Portal Keeper11
IA 101
Net
DNN
BLL - Entities
Extensions
Themes
IO
Services
Providers
Scheduler
Http Handlers
Pages Ashx Web API MVC
IISApplication
Http Context
Request
Response
Filters
Http Modules
Routing
DBFilesWeb Librairies
Portal Keeper
BotsAdapters Handlers Web
Services
Action
Agents
Firewall
AI Services + Demo 1
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Goal-based agents
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Building goal-based agents
bull What is the goal to be achieved
bull What are the actions
bull What is the states representation
Initial
state
Goal
state
Actions
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Example robotic assembly14
bull states coordinates of joint angles parts of the object to be assembled
bull actions continuous motions of robot joints
bull goal test complete assembly
bull path cost time to execute
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Example The 8-puzzle15
bull states locations of tiles
bull actions move blank left right up down
bull goal test = goal state (given)
bull path cost 1 per move
[Note optimal solution is NP-hard]
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Search
bull Uninformed Search
bull Informed Search ndash Heuristics
bull Adversarial Search ndash Games
bull Constraint Satisfaction problems
16
PathFindingjs SearchDemo 2
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Local Search
bull Hill Climbing Gradient descent
bull Problem depending on initial state can get stuck in local maxima
bull Simulated annealing
bull Stochastic Beam Search
bull Genetic algorithms
bull Genetic Sharp
17
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
A Model-Based Agent
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
A grammar of sentences in propositional logic
Parsing
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Proving things
bull A proof is a sequence of sentences derived by rules of inference
bull Last sentence = theorem (goal or query) to prove
bull Examples
1 Humid Premise ldquoIt is humidrdquo
2 HumidHot Premise ldquoIf it is humid it is hotrdquo
3 Hot Modus Ponens(12) ldquoIt is hotrdquo
bull But lacks expressivity
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
First Order Logic and KR languages
Logical inferenceDemo 3
Propositional Logic
First Order
Higher Order
Modal
Fuzzy
Logic
Multi-valued
Logic
Probabilistic
Logic
Temporal Non-monotonic
Logic
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Semantic Web
bull Resource Description Framework
bull KR community AAAI W3C Berners-Lee
bull RDF - triples (facts) class subclass
bull RDFS - OWL - defined classes constraints
bull SPARQL ndash Querying Triple Stores
bull Linked-Data - SOA
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
But real world is uncertain
bull Uncertain inputs
bull Missing noisy data
bull Uncertain knowledge
bull Multiple Incomplete conditions causes effects
bull Probabilisticstochastic effects
bull Uncertain outputs
bull Abduction and induction default reasoning
bull Incomplete inference
Probabilistic reasoning
Probabilistic results
23
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
24 Decision making with uncertainty
bull Rational behavior becomes
bull For each possible action identify the possible outcomes
bull Compute the probability of each outcome
bull Compute the utility of each outcome
bull Compute the expected utility over possible outcomes for each action
bull Select the action with Maximum Expected Utility
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Utility-based agents
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Probabilistic programming
bull Naiumlve Bayes model
119875 119862119886119906119904119890 1198641198911198911198901198881199051 hellip 119864119891119891119890119888119905119899
=119875 119862119886119906119904119890 lowast ς119894 119875 119864119891119891119890119888119905119894 119862119886119906119904119890
bull Bayesian networks
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Dynamic Bayesian Networks
bull Google 10 PageRank over a web graph
bull Transitionsbull With prob 1-c follow a random outlink (solid lines)
bull Stationary distribution bull Will spend more time on highly reachable pages
bull Markov processes
bull Hidden Markov Networksbull Hidden state observed evidence
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Example A simple weather Hidden Markov Model
bull An Hidden Markov Model is defined by
bull Initial distribution 119875(1198831)
bull Transitions 119875 119883119905 119883119905minus1)
bull Emissions 119875 119864 119883)
Probabilistic inferenceDemo 4
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Real Hidden Markov Models Examples
bull Natural Language processingbull Text classificationbull Information retrievalbull Information extraction
bull Speech recognition HMMs bull Observations are acoustic signalsbull States are positions in words
bull Machine translation HMMs bull Observations are wordsbull States are translation options
bull Radar tracking bull Observations are range readingsbull States are positions on a map
bull Trained probabilistic networks bull Given training set Dbull Find H that best matches D
bull Powerful toolkit InferNetbull Click-through
bull Sentiment analysis
Inducer
C
A
EB
][][][][
]1[]1[]1[]1[
MCMAMBME
CABE
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Different Learning tasks
bull Find ideal customersbull Amex
bull Find best person for jobbull BellAtlantic
bull Predict purchasing patternsbull Victoria Secret
bull Help win gamesbull NBA
bull Catalogue natural objectsbull Quasars
bull Bioinformaticsbull Identifying genes
bull Predicting protein function
bull Recognizing Handwritingbull Bell Labs ndash Lenet US Postal
bull Recognizing Spoken Wordsbull Ticketmaster
bull Translationbull Google
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Major paradigms of machine learning
bull Rote learning ndashAssociation-based storage and retrieval
bull Induction ndash Use specific examples to reach general conclusions
bull Clustering ndash Unsupervised identification of natural groups in data
bull Analogy ndash Correspondence between different representations
bull Discovery ndash Unsupervised specific goal not given
bull Reinforcement ndash Feedback at the end of a sequence of steps
31
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Learning agents
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
bull Construct consistent hypothesis to agree on training set
bull Example Regression curve fitting
bull Ockhamrsquos razor prefer the simplest hypothesis consistent with data
Inductive learning method
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Classification
bull Using higher dimensionsbull Linear classifier
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Artificial Neural Networks
bull Biological inspiration
bull Artificial Unit
bull Multiple layers
bull Expressiveness
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Deep learning
bull Deep networksbull Traditional MultiLayer classifier
bull Deep Natural Hierarchies
bull Convoluted Networks
bull Tensor Kernels
bull Subsampling
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Example Go
bull Game of GO
bull Simple but complex
bull Computer Go
bull 201603 - Alpha Go vs Lee Sedol
bull Deep learning Toolkits
bull Computing value Maps
Train and run a convoluted networkDemo 5
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you
DNNConnect2016
Questions
Please remember to evaluate the session online
Thank you