Upload
flytxt
View
612
Download
0
Tags:
Embed Size (px)
Citation preview
1confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Architecting IntelligenceBig Data Analytics and
Building Intelligent Applications
2confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Our discussion today
Big Data Analytics and Intelligent Applications• The puzzle, the hype, the customer?
• Man-machine collaboration
State of the Art• Practical AI, Machine Learning, Data Mining
• Data Science, Data Games
What does the future guarantee?• Physics, Networks and Computation
• New computation models?
3confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Big Data, the Meme
4confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Big (Data Analytics) Distraction
5confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Big (Data) Crowd
6confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Start unraveling the complexity
What do I want to communicate that currently requires a significant amount of time and energy to analyze, interpret, and share?• Stuart Frankel, “Data Scientists Don’t Scale”, Harvard Business Review,
May 2015
What economic value will my customer gain from Big Data?
7confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
FlytxtOur vision is to create >10% measurable economic value for Mobile Enterprises through Big Data Analytics
Flytxt’s solutions create incremental revenues from new and existing sources, optimize margins and enhance customer experience
Dutch company with corporate office in Dubai, global delivery centres in India and regional presence in Mexico City, Johannesburg, Singapore, Dhaka and Nairobi.
Sample text
Awards and Recognitions
8confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Benefits delivered to customers
PartnersOperators
Proven across many Countries, Brands and Logos
Brands
IIT DELHI
4% Increase
in Gross Revenue
30%Growth
in Mobile Money users
10%Growth
in Data Users
105%Increase
in Special offer Sales
300%Increase
in Store Footfall
25%Dropin Churn
9confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Big Data Technology Architecture – the Flytxt example
10confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Our discussion today (Part 2)
Big Data Analytics and Intelligent Applications• The puzzle, the hype, the customer?
• Man-machine collaboration
State of the Art• AI, Machine Learning, Data Mining
• Data Science, Data Games
What does the future guarantee?• Physics, Networks and Computation
• New computation models?
11confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Practical AI: Personalized Driving
Navigator suggest alternative route due to traffic congestion
System identifies primary/secondary driver
Bluetooth connection to Car
Systems
Car system connects with
database to access unique ID info
Infotainment settings are
modified (language
preference, radio station…)
Navigator presents favorite destinations
Connected office identifies next meeting happens in 10 min and offers re-scheduling
12confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Practical AI: Pattern Classification in Location Analytics
6/30/2Copyright © 2012 12
Automatic classification of venues / routesbased on their features
Each venue/route is represented by a setof features
Labeled examples corresponding tovarious venue types / route types whichrepresent classes
Learn a decision boundary that separatesthe classes & then make predictions
13confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Customer MarketingProgram Product Rol
Infra-structure
Location
Descriptive
Exploratory
Heuristic
Predictive
Prescriptive
Visualization
CLV Monitoring
Opportunity Identification
Behavioral Variations
Action Prediction
Personalized Recommendation
Effectiveness Measurement
Program Reach Analysis
Business impact Analysis
Outcome Forecasting
Impact Optimization
Product Popularity Monitoring
Product Promotion Analysis
Product Association
Profitability Simulation
Product Promotion Recommendation
Business Health Monitoring
KPI Impact Analysis
Business Impact measurement
Impact Forecasting
Yield Optimization
Utilization Monitoring
Challenge Identification
Cost Benefit Analysis
Event Prediction
Optimization Recommendation
Geo-Spatial Reporting
Location Affinity Analysis
Location- Behavior Association
Location based Forecasting
Location based Recommendation
Roots of Practical AI: Analytics Models built by Data Scientists
14confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Data Sciences: State of the Art
KDD Cup: organized by ACM Special Interest Group on Knowledge Discovery and Data Mining
2010: Predict student performance on mathematical problems from Intelligent Tutoring System logs
2011: Recommending Music Items based on the Yahoo! Music Dataset
2012: Predict which information sources one user might follow in Weibo (Chinese “twitter”)
2013: Determine whether an author has written a given paper
2014: Predict funding requests that deserve an A+ (for DonorsChoose.org)
2015: Predict student dropout on a Massive Open Online Course platform (XuetangX)
15confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Our discussion today (Part 3)
Big Data Analytics and Intelligent Applications• The puzzle, the hype, the customer?
• Man-machine collaboration
State of the Art?• AI, Machine Learning, Data Mining
• Data Science, Data Games
What does the future guarantee?• Physics, Networks and Computation
• New computation models?
16confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Memoirs from the Past: Hilbert’s Program
In 1900, David Hilbert, a very influential universal mathematician, announced a grand search for a complete and consistent set of axioms for all mathematics
In 1931, Kurt Gödel announced his discovery of the Incompleteness Theorem: There will always be statements about the natural numbers that are true, but that are unprovable within the system
Hilbert probably dedicated his life trying to prove his hypothesis, which Gödel proved cannot be true!
However, Gödel’s work inspired Alan Turing and Alonzo Church, and in 1936, they mathematically defined “computation”
17confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
The future AI platform is a network!
Courtesy: Maulik Kamdar, Stanford University
18confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Future AI agents
AI agents will compute; with data that gets generated on many devices
2025: 100 billion connected devices, 175 zeta bytes of data per year (Huawei)
Data volumes will grow faster than any network or computer can be sized
How will you scale the AI of tomorrow?
19confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Practical AI: Moving data, moving code
Code must meet data to compute – code moves and/or data does, across a (wireless) network
History: All data moved to where the code was
Near past: Parallel and distributed computation – partition code & data
Now: (approximately) Move code to where the data is (Hadoop etc)
Future: Determine the code-data match and optimize movement?• Is there is a computational model for this?
20confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Physics, Networks, Computation – immutable laws
Energy dissipation in radiation (Gauss’s / Coulomb’s Laws)• Low energy reception implies higher decoding error (Shannon’s Limit)• How fast can memory-to-memory transfers happen?
Capacity of a wireless network is constrained by interference (e.g. see Gupta & Kumar, 2000)• Spectrum (# channels) available will remain finite• Channel allocations will be dynamic, but how fast can two interfering pairs
find free channels?
Are there limits to local computation? (e.g. see works by Ning Xie, Shai Vardi)• Moving code or data implies “local” processing• How much AI can be computed, and at what cost?
21confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Example: High-dimension Clustering
6/30/2 Co21
Basic machine learning algorithm to group nodes (users, people, devices) by state (behavior)• Each node produces a vector describing current state
• Nodes are clustered together by some measure of vector similarity
“Moving code” distributed implementations available today (on Hadoop/Spark)
Future: Rate of change of state will outpace speeds of computation and communication
Is the solution hierarchical, is the paradigm divide and conquer?
How will network & algorithm design and implementation change?• Can all clustering problems be solved “locally”?
22confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Discussion summary
Big Data Analytics and Intelligent Applications• Build for customer value, build simple solutions
State of the Art• Practical AI and Data Sciences
What does the future guarantee?• Need to scale AI compute: Data generation rates faster than compute
/ communication rates
23confidentialFlytxt. All rights reserved. 30 June 201530 June 2015©
Thank Youwww.flytxt.com