Artificial Intelligence Strategy: For Small Business

Georgia Mentor Protégé Connection
October 18, 2017

Transcript

1. CHRIS BENSON ARTIFICIAL INTELLIGENCE STRATEGY FOR SMALL BUSINESS
2. CHRIS BENSON ▸ Chief Scientist for Artificial Intelligence & Machine Learning
 
 
 
 Safety & Productivity Solutions ▸ Formerly: Artificial Intelligence Delivery Manager
 
 ▸ Artificial Intelligence & Machine Learning Strategist ▸ Deep Learning Architect (hands-on engineering) ▸ Pubic Speaker on Artificial Intelligence topics ▸ Organizer, Atlanta Deep Learning Meetup ▸ Introduced to Deep Learning in 1992 ▸ Futurist • Productive Disruptor • Optimist with Reservations ▸ First ‘Hello World’ using IBM BASIC at age 11 on this…
5. INTRODUCED TO DEEP LEARNING 25 YEARS AGO
6. F-22 RAPTOR WORLD’S MOST ADVANCED AIR-SUPERIORITY STEALTH FIGHTER LOCKHEED MARTIN DESIGNED, DEVELOPED, AND BUILT AT THE LOCKHEED MARTIN PLANT NEAR ATLANTA
7. YF-22 CRASH LANDING EDWARDS AFB APRIL 1992 AN AVIONICS ERROR FAILED TO PREVENT A PILOT-INDUCED OSCILLATION.
8. THIS CRITICAL YF-22 AVIONICS ERROR WAS SOLVED (IN PART) BY A VETERAN LOCKHEED ENGINEER WHO HAD THE INSPIRATION TO RESEARCH AND APPLY A BLEEDING-EDGE DEEP LEARNING APPROACH TO THE AVIONICS PROBLEM. HIS NAME WAS WHIT BENSON, AND HE WAS MY FATHER.
 (1932 - 2011)
9. HISTORY OF ARTIFICIAL INTELLIGENCE THE TWO ERAS
10. SYMBOLIC AI HIGH-LEVEL “SYMBOLIC" (HUMAN-READABLE) REPRESENTATIONS OF PROBLEMS, LOGIC, SEARCH 1950S - 1980S COGNITIVE SIMULATION AND LOGIC-BASED ( E.G. ABSTRACT REASONING, PROBLEM-SOLVING, SIMULATE 'HUMAN COGNITION’ ) KNOWLEDGE-BASED ( E.G. EXPERT SYSTEMS ) SOURCE: HTTPS://EN.WIKIPEDIA.ORG/WIKI/ARTIFICIAL_INTELLIGENCE
11. STUDY AND CONSTRUCTION OF ALGORITHMS THAT CAN LEARN FROM AND MAKE PREDICTIONS ON DATA 1980S - PRESENT DAY GIVES COMPUTERS THE ABILITY TO LEARN WITHOUT BEING EXPLICITLY PROGRAMMED BUZZWORDS EXTRAORDINAIRE! SOURCE: HTTPS://EN.WIKIPEDIA.ORG/WIKI/MACHINE_LEARNING MACHINE LEARNING / DEEP LEARNING
12. SOURCE: HTTPS://BLOG.UBISEND.COM/DISCOVER-CHATBOTS/HOW-TO-MAKE-A-CHATBOT-THAT-LEARNS
13. MACHINE LEARNING / DEEP LEARNING SOURCE: HTTPS://WWW.COURSERA.ORG/LEARN/MACHINE-LEARNING ▸ Linear Regression ▸ Logistic Regression ▸ Support Vector Machines ▸ K-Means Clustering ▸ Principal Components Analysis ▸ Anomaly Detection ▸ Recommender Systems ▸ Collaborative Filtering ▸ Supervised vs Unsupervised Learning ▸ Dimensionality Reduction ▸ Regularization & Optimization ▸ Large-Scale ML System Design ▸ Deep Learning / Neural Networks ▸ Machine Perception / Computer Vision / Object Recognition ▸ Speech Recognition / Natural Language Processing
14. WHAT IS DEEP LEARNING? THE COCKTAIL
 PARTY ANSWER
15. DEEP LEARNING IS MODERN ARTIFICIAL INTELLIGENCE THAT WORKS TODAY
16. DEEP LEARNING IS THE APPROACH TO MACHINE LEARNING THAT IS DRIVING THE CURRENT ARTIFICIAL INTELLIGENCE REVOLUTION
17. LEARNING FROM DATA WITHOUT EXPLICIT PROGRAMMING
18. DEEP LEARNING IS THE MODERN APPLICATION OF DEEP NEURAL NETWORKS TO ACHIEVE MACHINE LEARNING
19. DEEP LEARNING IN CONTEXT HOW THEY RELATE ARTIFICIAL INTELLIGENCE MACHINE LEARNING DEEP LEARNING
20. DEEP LEARNING IN CONTEXT JUXTAPOSITION OF FOUR DISCIPLINES DATA SCIENCE ARTIFICIAL INTELLIGENCE BIG DATA MACHINE LEARNING DEEP LEARNING
21. “IN RECENT YEARS, IT (DEEP LEARNING) HAS SEEN TREMENDOUS GROWTH IN ITS POPULARITY AND USEFULNESS, DUE IN LARGE PART TO MORE POWERFUL COMPUTERS, LARGER DATASETS AND TECHNIQUES TO TRAIN DEEPER NETWORKS.” Ian Goodfellow, Yoshua Bengio & Aaron Courville (2016). Deep Learning. MIT Press WHAT IS DEEP LEARNING
22. WHY DOES DEEP LEARNING MATTER?
23. DEEP NEURAL NETWORKS CAN PRECISELY APPROXIMATE ANY CONTINUOUS FUNCTION
24. IF INPUT, THEN OUTPUT
25. AND NEARLY ANY PROCESS IMAGINABLE INVOLVES COMPUTING A FUNCTION
26. THIS IS HUGE NEURAL NETWORKS ARE
 UNIVERSAL
27. DEEP LEARNING IS A
 UNIVERSAL
 APPROACH TO SOLVING COMPLEX PROBLEMS
28. “THAT’S COOL, CHRIS…
 BUT I’M A BUSINESS PERSON NOT A DATA SCIENTIST WHY SHOULD I CARE ABOUT
 DEEP LEARNING?”
29. “THE LAST 10 YEARS HAVE BEEN ABOUT BUILDING A WORLD THAT IS MOBILE-FIRST… BUT IN THE NEXT 10 YEARS, WE WILL SHIFT TO A WORLD THAT IS AI-FIRST…” AI-FIRST GOOGLE CEO SUNDAR PICHAI A PERSONAL GOOGLE, JUST FOR YOU SOURCE: HTTPS://WWW.BLOG.GOOGLE/PRODUCTS/ASSISTANT/PERSONAL-GOOGLE-JUST-YOU
30. “IN AN AI-FIRST WORLD, WE ARE RETHINKING ALL OUR PRODUCTS AND APPLYING MACHINE LEARNING AND AI TO SOLVE USER PROBLEMS. AND WE ARE DOING THIS ACROSS EVERY ONE OF OUR PRODUCTS.” AI-FIRST GOOGLE CEO SUNDAR PICHAI’S KEYNOTE AT 2017 I/O CONFERENCE SOURCE: HTTPS://SINGJUPOST.COM/GOOGLE-CEO-SUNDAR-PICHAIS-KEYNOTE-AT-2017-IO-CONFERENCE-FULL-TRANSCRIPT
31. ARTIFICIAL INTELLIGENCE IS NO LONGER JUST FOR BIG TECH COMPANIES
32. EVERY INDUSTRY ON THE PLANET WILL BE IMPACTED BY ARTIFICIAL INTELLIGENCE
33. "ARTIFICIAL INTELLIGENCE IS THE FUTURE, NOT ONLY OF RUSSIA, BUT OF ALL OF MANKIND.” "THE INDUSTRY LEADER WILL RULE THE WORLD.” - PUTIN, SEPTEMBER 1, 2017
35. IN THE YEARS TO COME, YOU WILL CERTAINLY USE, AND MAYBE EVEN CREATE, ARTIFICIAL INTELLIGENCE PRODUCTS AND SERVICES. MANY OF THOSE WILL BE BASED ON DEEP LEARNING
36. “IF CIOS INVESTED IN MACHINE LEARNING THREE YEARS AGO, THEY WOULD HAVE WASTED THEIR MONEY. BUT IF THEY WAIT ANOTHER THREE YEARS, THEY WILL NEVER CATCH UP.” DAN OLLEY, ELSEVIER CTO CIO MAGAZINE - APRIL 26, 2016 WHY IT’S TIME FOR CIOS TO INVEST IN MACHINE LEARNING SOURCE: HTTP://WWW.CIO.COM/ARTICLE/3061713/LEADERSHIP-MANAGEMENT/WHY-ITS-TIME-FOR-CIOS-TO-INVEST-IN-MACHINE-LEARNING.HTML
38. “THE YEARS AHEAD ARE FULL OF CHALLENGES AND OPPORTUNITIES TO IMPROVE DEEP LEARNING EVEN FURTHER AND BRING IT TO NEW FRONTIERS.” Ian Goodfellow, Yoshua Bengio & Aaron Courville (2016). Deep Learning. MIT Press THE FUTURE OF ARTIFICIAL INTELLIGENCE
39. WHAT IS DEEP LEARNING? THE MAD SCIENTIST ANSWER
40. “DEEP LEARNING IS AN APPROACH TO MACHINE LEARNING THAT HAS DRAWN HEAVILY ON OUR KNOWLEDGE OF THE HUMAN BRAIN, STATISTICS AND APPLIED MATH.” Ian Goodfellow, Yoshua Bengio & Aaron Courville (2016). Deep Learning. MIT Press WHAT IS DEEP LEARNING
41. “DEEP LEARNING IS AN APPROACH TO AI. SPECIFICALLY, IT IS A TYPE OF MACHINE LEARNING, A TECHNIQUE THAT ALLOWS COMPUTER SYSTEMS TO IMPROVE WITH EXPERIENCE AND DATA.” Ian Goodfellow, Yoshua Bengio & Aaron Courville (2016). Deep Learning. MIT Press WHAT IS DEEP LEARNING
42. “BY GATHERING KNOWLEDGE FROM EXPERIENCE, THIS APPROACH AVOIDS THE NEED FOR HUMAN OPERATORS TO FORMALLY SPECIFY ALL THE KNOWLEDGE THAT THE COMPUTER NEEDS.” Ian Goodfellow, Yoshua Bengio & Aaron Courville (2016). Deep Learning. MIT Press WHAT IS DEEP LEARNING
43. “IN RECENT YEARS, IT (DEEP LEARNING) HAS SEEN TREMENDOUS GROWTH IN ITS POPULARITY AND USEFULNESS, DUE IN LARGE PART TO MORE POWERFUL COMPUTERS, LARGER DATASETS AND TECHNIQUES TO TRAIN DEEPER NETWORKS.” Ian Goodfellow, Yoshua Bengio & Aaron Courville (2016). Deep Learning. MIT Press WHAT IS DEEP LEARNING
45. HOW DOES
 DEEP LEARNING WORK?
47. DEEP LEARNING ARCHITECTURES CONVOLUTIONAL NETWORKS FEED-FORWARD CONNECTIVITY INSPIRED BY THE ANIMAL VISUAL CORTEX. VISUAL 'TILING' ENABLES IMAGE AND VIDEO RECOGNITION, RECOMMENDER SYSTEMS, AND NATURAL LANGUAGE PROCESSING. RECURRENT NETWORKS CONNECTIONS CREATE AN INTERNAL MEMORY FOR DYNAMIC TEMPORAL BEHAVIOR LIKE SPEECH RECOGNITION OR HANDWRITING RECOGNITION. GENERATIVE ADVERSARIAL NETWORKS TWO NEURAL NETWORKS COMPETING AGAINST EACH OTHER - ONE GENERATIVE AND ONE DISCRIMINATIVE. BLEEDING-EDGE APPROACH USING UNSUPERVISED TRAINING. FEED-FORWARD NETWORKS WITH BACKPROPAGATION THIS IS THE ORIGINAL AND MOST COMMON FORM OF DEEP NEURAL NETWORK.
 WE WILL EXPLORE BACKPROPAGATION IN THE SLIDES TO COME.
48. TRAINING ENABLES GENERALIZED PREDICTIONS USING KNOWN CORRECT RESULTS
49. LOTS OF KNOWN CORRECT RESULTS
50. ONCE TRAINED,
 GENERALIZED PREDICTIONS FROM NEW INPUTS
51. HOW DEEP LEARNING WORKS - FEED FORWARD NEURAL NETWORK INPUT INPUT INPUT OUTPUT DATA DATA DATA In training, data flows forward through the layers of the neural network.
52. EACH NEURON SOURCE: HTTPS://THECLEVERMACHINE.WORDPRESS.COM/TAG/BACKPROPAGATION
53. HOW DEEP LEARNING WORKS - ERROR BACKPROPAGATION OUTPUT DATA DATA DATA CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS CALC ERROR & ADJUST WEIGHTS EXPECTED VS INPUT INPUT INPUT The actual output values are compared to the known target values of the training data. Then moving backward from output to input, each node’s error is used to adjust that node’s level of importance, in the hope that the next iteration will be more accurate.
54. HOW DEEP LEARNING WORKS DATA OUTPUT INPUT INPUT INPUT DATA DATA Keep iterating forwards and backwards until the neural network’s accuracy meets your goals.
55. “UPDATE EACH OF THE WEIGHTS IN THE NETWORK SO THAT THEY CAUSE THE ACTUAL OUTPUT TO BE CLOSER THE TARGET OUTPUT, THEREBY MINIMIZING THE ERROR FOR EACH OUTPUT NEURON AND THE NETWORK AS A WHOLE.” BACKPROPAGATION SOURCE: HTTPS://MATTMAZUR.COM/2015/03/17/A-STEP-BY-STEP-BACKPROPAGATION-EXAMPLE
56. “BASICALLY, TRAINING IS A SEARCH. YOU ARE SEARCHING FOR THE SET OF WEIGHTS THAT WILL CAUSE THE NEURAL NETWORK TO HAVE THE LOWEST GLOBAL ERROR FOR A TRAINING SET.” Jeff Heaton (2012). Introduction to the Math of Neural Networks. Heaton Research, Inc. BACKPROPAGATION
57. DEEP LEARNING IN PRODUCTION
59. https://medium.com/chat-bots-weekly/chatbot-conversational-ui-start-here-2f9250e8cde0
60. COMMON DEEP LEARNING USE CASES ▸ Anomaly Detection / Cyber-Security / Fraud Detection ▸ Recommender Systems / Marketing Personalization / Search ▸ Machine Perception / Computer Vision / Object Recognition ▸ Speech Recognition / Natural Language Processing ▸ Transportation (e.g. self-driving cars) / Drone Navigation ▸ Healthcare Diagnosis / Medical Imaging Interpretation ▸ Securities Trading / Financial Analysis / Economic Forecasting
61. QUESTIONS TO START WITH ▸ What problem are we trying to solve? ▸ What data do we need to solve the problem? ▸ What data do we have? Where is it coming from? ▸ Do we have enough of the right data to train models? ▸ Where will that data be aggregated and maintained? ▸ What pre-processing / post-processing will be necessary? ▸ What enterprise and data architectures will be used?
62. MACHINE LEARNING VENDORS WITH MANY MORE ENTERING THE MARKETPLACE
64. THE FUTURE OF ARTIFICIAL INTELLIGENCE TENSORFLOW FROM GOOGLE IS THE MARKET LEADER
65. THE FUTURE OF ARTIFICIAL INTELLIGENCE WHILE IBM WATSON HAS STAKED OUT THE ENTERPRISE
67. ▸Twitter: @chrisbenson
 ▸LinkedIn: https://linkedin/in/chrisbenson
 I invite you to connect with me.
 ▸Atlanta Deep Learning Meetup
 http://atlantadeeplearning.org
 ▸Speaker Deck for this presentation
 https://speakerdeck.com/benson/artificial- intelligence-strategy-for-small-business THANK YOU VERY MUCH!