Has AI Arrived? by Paco Nathan

https://www.youtube.com/watch?v=v0t7ZO8Ah4A&list=PL6O3g23-p8Tr5eqnIIPdBD_8eE5JBDBik&index=2&t=470s

Transcript

2. Has AI Arrived? Big Data Spain
 Madrid, 2016-11-17 Paco Nathan, @pacoid
 Director, Learning Group @ O’Reilly Media 1
3. A rhetorical question: from:
 Beyond the AI Winter
 goo.gl/tKug8u Can you name ten successful tech start-ups which lack 
 any application of Machine Learning on their roadmaps? 2
4. An interesting perspective: To paraphrase Peter Norvig, Google @ AI Conference 2016: Marc Andreessen noted famously how software
 was disrupting so many incumbents … and now 
 Machine Learning is disrupting many incumbents from:
 Software engineering of systems that learn in uncertain domains
 safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260721.html 3
5. A related perspective: Pedro Domingos believes we’re getting closer to realizing
 a “universal learner” The future belongs to those who understand at
 a very deep level how to combine their unique
 expertise with what algorithms do best. from:
 The Master Algorithm
 goodreads.com/book/show/24612233-the-master-algorithm 4
6. A related perspective: Domingos describes “five tribes” of machine learning 
 (see especially on page 54): • symbolists: inverse deduction, e.g., rule systems • connectionists: what the brain does, e.g., deep learning • evolutionaries: natural selection, e.g., genetic programming • bayesians: uncertainty, e.g., probabilistic inference • analogizers: similarities, e.g., support vectors from:
 The Master Algorithm
 goodreads.com/book/show/24612233-the-master-algorithm 5
7. In retrospect: During the past few years applications of deep learning have exploded. Among those tribes, “connectionists” now prevail. Even so, deep learning is only a portion of machine learning. Moreover machine learning does not represent the entirety 
 of machine intelligence. What else will be needed? 6
8. Where are the examples? 7
9. Major tech firms (just a sample): 8
10. “An Ecosystem of Machine Intelligence” oreilly.com/ideas/the-current-state-of-machine-intelligence-3-0 Shivon Zilis, James Cham, Heidi Skinner 9
11. Reaching Human Parity: Historic Achievement: Microsoft researchers reach human parity 
 in conversational speech recognition blogs.microsoft.com/next/2016/10/18/historic-achievement-microsoft- researchers-reach-human-parity-conversational-speech-recognition/ 10
12. Reaching Human Parity: Historic Achievement: Microsoft researchers reach human parity in conversational speech recognition blogs.microsoft.com/next/2016/10/18/historic-achievement-microsoft- researchers-reach-human-parity-conversational-speech-recognition/ Shades of HAL: openreview.net/pdf? id=BkjLkSqxg 11
13. Realistically… consider the control system at the heart of, say, Uber – 
 manipulating supply chains of resources for particular outcomes 12
14. Some favorite examples in arts & lit: Benjamin.ai / Sunspring youtu.be/LY7x2Ihqjmc 13
15. Some favorite examples in arts & lit: Flash Forward: “The Witch Who Came From Mars” flashforwardpod.com/2016/09/05/episode-20-something-martian-witch-way-comes/ 14
16. Artificial Intelligence conference series: New York City (last Sep) conferences.oreilly.com/artificial-intelligence/ai-ny-2016 San Francisco (last Oct) conferences.oreilly.com/artificial-intelligence/bot-ca New York City, Jun 26-29 2017 conferences.oreilly.com/artificial-intelligence/ai-ny (CFP open through Jan 18) 15
17. Artificial Intelligence conference series: New York City (last Sep) conferences.oreilly.com/artificial-intelligence/ai-ny-2016 San Francisco (last Oct) conferences.oreilly.com/artificial-intelligence/bot-ca New York City, Jun 26-29 conferences.oreilly.com/artificial-intelligence/ai-ny (CFP open through Jan 18) As one might imagine, the presenters discussed much 
 deep learning – although there were other important points… let’s consider those 16
18. AI requires sophisticated engineering? Software engineering of systems that learn in uncertain domains safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260721.html 17
19. Observations by Peter Norvig: • difficult to debug, revise incrementally, verify • less transparency into algorithms • components are hard to isolate, for debugging • automated integration introduces unusual risks • tech debt accumulates more readily Machine Learning: The High Interest Credit Card of Technical Debt research.google.com/pubs/pub43146.html Software engineering of systems that learn in uncertain domains safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260721.html AI requires sophisticated engineering? 18
20. Why should I trust you? Explaining the predictions of any classifier safaribooksonline.com/library/view/strata-hadoop/ 9781491944660/video282744.html kdd.org/kdd2016/subtopic/view/why-should-i-trust-you- explaining-the-predictions-of-any-classifier Carlos Guestrin: LIME 19
21. Impact on Big Data, Cloud, etc.: Overall, AI drives product features That process in turn drives cloud consumption 
 (look at the major players) What’s the impact for those already immersed 
 in Big Data, Data Science, Machine Learning, Distributed Systems, Cloud technologies, 
 DevOps practice, etc.? In word: Good The results will be in health
 care, manufacturing, agriculture,
 energy, transportation, etc. 20
22. Artificial intelligence: making a human connection safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260723.html AI work is mostly human? 21
23. Observations by Genevieve Bell @ Intel: An anthropologist would ask: “Who raised you? 
 Who were your mummies and your daddies?” ... 
 AI has had a lot of daddies. If we understand the founders, we can ask what 
 do we need to bring back into the conversation? Artificial intelligence: making a human connection safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260723.html AI work is mostly human? 22
24. AI work is mostly human? The Future of AI, Oren Etzioni @ AI2 safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video282377.html 23
25. Etzioni stressed the key role of humans-in-the-loop: 99% of machine learning is human work AI work is mostly human? 24
26. Over-anthropomorphization may become problematic: • does this analysis introduce unneeded bias? • machine intelligence differs from human cognition, 
 e.g., abductive reasoning (e.g., C.S. Peirce) • consider examples of evolved antenna AI work is mostly human? 25
27. Jobs won’t be displaced by AI? Why we’ll never run out of jobs safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260722.html 26
28. Observations by Tim O’Reilly: We won’t run out of work until we run out of problems Our main advances have come when we invested in other people's children – massive investment in EU following WWII, built from something that resembles Syria today 21st c great question: “Who’s black box do you trust?” Jobs won’t be displaced by AI? Why we’ll never run out of jobs safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260722.html 27
29. US voting by state g.co/kgs/PSq9JS Jobs won’t be displaced by AI? 28
30. US jobs by state npr.org/sections/money/2015/02/05/382664837/map-the-most-common-job-in-every-state Jobs won’t be displaced by AI? 29
31. Realistically, fully self-driving trucks are a bit further away fool.com/investing/2016/10/30/despite-ubers-self-driving-truck- delivery-truck-dr.aspx Some contend that no existing economic model addresses 
 the accelerating pull of technological deflation Meanwhile, social reforms regarding health care and
 Universal Basic Income become urgent priorities Jobs won’t be displaced by AI? 30
32. Does AI = Deep Learning? Obstacles to progress in AI safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260902.html 31
33. Yann LeCun described some necessary components of AI: • perception • predictive model • memory • reasoning and planning Obstacles to progress in AI safaribooksonline.com/library/view/oreilly-ai-conference/ 9781491973912/video260902.html Does AI = Deep Learning? 32
34. AI is much more than Deep Learning Perception, prediction, memory – these are necessary; however, they do not address understanding Winograd Schemas show the need for common sense and contextual understanding – replacement for Turing Test see: The Winograd Schema Challenge Hector Levesque commonsensereasoning.org/2011/papers/Levesque.pdf 33
35. AI is much more than Deep Learning Common sense and context: for example, without ample knowledge of the world, a sentence cannot be understood 㱺 embodied cognition (prevailed for a while) 㱺 ontology (more difficult, likely much more useful) 34
36. A lesson from history see: Why AM and Eurisko Appear to Work Doug Lenat, John Seely Brown aaaipress.org/Papers/AAAI/1983/AAAI83-059.pdf Eurisko, The Computer With A Mind Of Its Own
 George Johnson aliciapatterson.org/stories/eurisko-computer-mind-its-own Eurisko, and a mobius strip memory cell Learning, rules, patterns – these only go so far Ontology and the quest for common sense 35
37. Some Missing Pieces With ML, we assume there’s structure embedded in the 
 data, then build ML models to validate those assumptions However, which tools serve to identify structure? see: Persistent Homology: An Introduction and a New Text Representation 
 for Natural Language Processing Xiaojin Zhu pages.cs.wisc.edu/~jerryzhu/pub/homology.pdf Topological Data Analysis Chad Topaz dsweb.siam.org/TheMagazine/Article/TabId/823/ArtMID/1971/ArticleID/777/ Topological-Data-Analysis.aspx 36
38. AI transformations Recently launched our own AI project within O’Reilly Media… We’re not a high-tech company; even so, the value of our data gets unlocked through AI This project makes use of cloud, Spark, Mesos, Kubernetes, Docker, etc., leveraging the tools we know, but in more complex use cases now. 37
39. 13K lexemes: our “universe” for customer interaction Too much cognitive load for any editor or engineer to master; however, not so difficult for a small cluster. Curation is hard; you don’t want it full automated – related to what Norvig calls the “Inattention Valley” AI transformations 38
40. Challenge: generating an implicit graph versus curating 
 an explicit graph, then maintaining integrity between: A C B E D ML, Big Data, etc.: computed similarity, inferred links, etc. (empiricists) Curated ontology: graph queries, 
 search rewrites, etc. (rationalists) a c b e d AI transformations 39
41. A C B E D a c b e d Needs better tooling 
 (SPARQL and triple store crowd haven’t gotten the memo 
 yet about containers, orchestration, microservices, etc.) AI transformations BTW, this repo is fantastic: github.com/danielricks/penseur 40
42. David Beyer: Reshaping global industries Machine intelligence in the wild: How AI will reshape global industries safaribooksonline.com/library/view/strata-hadoop/9781491944660/ video282803.html 41
43. To paraphrase: Consider the shift from steam to electric power: it took a generation before factory managers understood they could reconfigure the physical arrangement AI may be quicker adoption, but faces similar extremes of cognitive embrace Machine intelligence in the wild: How AI will reshape global industries safaribooksonline.com/library/view/strata-hadoop/9781491944660/ video282803.html David Beyer: Reshaping global industries 42
44. Looking ahead… We have a need now to distinguish between what humans and computers can do well, respectively cognitive load, speed, scale, repeatability:
 computers > humans curation (captchas, as an example):
 computers < humans Organizations which focus on this 
 expertise for AI applications will
 have a distinct advantage 43
45. presenter: Just Enough Math O’Reilly (2014) justenoughmath.com monthly newsletter for updates, 
 events, conf summaries, etc.: liber118.com/pxn/
 @pacoid