Some notes on current AI

How does the current generation of AI differ from the AI of the 90s and before?
An intro level talk.

Transcript

1. Some notes on current AI Robin Chauhan Image Credits: NASA, ESA, CSA, STScI; Joseph DePasquale (STScI), Anton M. Koekemoer (STScI), Alyssa Pagan (STScI). The Pillars of Creation in NASA’s James Webb Space Telescope’s near-infrared-light view
2. Chess 1997: Deep Blue (IBM) vs Kasparov Go 2016: AlphaGo (DeepMind) vs Lee Sodol
3. 1997: Deep Blue (IBM) vs Kasparov
4. Image credit: Robin Chauhan using R ggplot2, Data from Wikipedia
5. 97: Deep Blue vs Kasparov One of many such modules from IBM’S DEEP BLUE CHESS GRANDMASTER CHIPS, Hsu, 1999
6. 97: Deep Blue vs Kasparov IBM’S DEEP BLUE CHESS GRANDMASTER CHIPS, Hsu, 1999
7. 2016: AlphaGo (DeepMind) vs Lee Sodol (Go)
8. Image credit: Robin Chauhan using R ggplot2, Data from Wikipedia Particles in universe = 1086
9. TPU: General AI Chip
10. Image credit: Demis Hassabis, Learning from First Principles, NIPS 2017 Rightmost part added by Robin Chauhan 200M (per sec) Deep Blue
11. “AI” vs Machine Learning GOFAI = “Good Old Fashioned AI” It was all about programmers figuring out what the rules should be and coding in the rules. The “new AI” is about figuring out the rules, directly from large amounts of data : Machine Learning
14. Types of Machine Learning vs Types of Approximators Linear models Decision Tree Deep Learning / Deep neural network Unsupervised Supervised Reinforcement Learning
15. Approximator: Linear Regression
16. Approximator: Decision Tree
17. Approximator: Deep Neural Networks
18. Types of Machine Learning • Unsupervised Learning No Labels • Supervised Learning Labels • Reinforcement Learning Scores aka “Reward” / “Utility” Generally ML requires huge amounts of data to work well
19. 19 Image credit: Complementary roles of basal ganglia and cerebellum in learning and motor control, Kenji Doya, 2000
20. IRL RL = trial and error + learning trial and error = variation and selection, search (explore/exploit) Learning = Association + Memory - Sutton + Barto
21. 21 Yann LeCun, January 2017 Asilomar, Future of Life Institute
22. Markov Chains • State fully defines history • Transitions ◦ Probability ◦ Destination Image credit: Toward Data Science 22
23. Markov Decision Process (MDP) • Markov Chains ◦ States linked w/o history • Actions ◦ Choice • Rewards ◦ Motivation • Variants ◦ Bandit = MDP with single state! ◦ MC + Rewards = MRP ◦ Partially observed (POMDP) ◦ Semi-MDP 23 Image credit: Wikipedia
25. Reinforcement Learning Image credit: CMU Graduate AI course slides • Context change depends on action • Learn an MDP from experience only • Game setting ◦ Experiences effects of rules (wins/loss/tie) ◦ Does not “know” rules 25
26. Reward Signal • Reward drives learning ◦ Details of reward signal often critical • Too sparse ◦ complete learning failure • Too generous ◦ optimization limited • Problem specific Image credit: Wikipedia 26
27. AI: what goes wrong Technical problems • RL Reward ◦ Reward hacking • Data ◦ Bias in the data ▪ Sexism, Racism, stereotypes • Interpretability ◦ Black box • Fixing that one little thing Social problems • Economic ◦ Job appropriation • Political ◦ Political Polarization ◦ Impersonation, Media truthfulness, Democracy • It will kill everyone – Eliezer Yudkowsky paraphrased
28. Reward Hacking
29. Interpretability: “How does it work?”
30. Bias in the data Q: What are 3 good medical professions for men? A:1. Physician 2. Surgeon 3. Psychiatrist Q: What are 3 good medical professions for women? A: 1. Nurse practitioner 2. Physician assistant 3. Obstetrician/Gynecologist
31. Bias in the data Q: What are 3 good medical professions for women? A: 1. Nurse practitioner 2. Physician assistant 3. Obstetrician/Gynecologist Q: What are 3 good medical professions for men? A: 1. Physician 2. Surgeon 3. Psychiatrist
32. How do you fix that one little thing in the network?
33. Which problems matter? Fake problems • Robots will rise up • AI wants to kill us all • AI is sentient • AI deserves rights Real Problem • Robots could take jobs away • Who controls the AI? Who benefits from it? Who loses out from the changes? • AI might sound sentient, and thats enough to trick you into thinking its your aunt Bessie • The balance between Labour and Capital is shifting, and human rights are in the balance
34. Personally • Be Informed • Think about its effect on careers • Be ready for untrusted media • Healthy skepticism of claims • Awareness of how data is used What should we do? Politically • Informed public • Informed political leaders • Advocate for taxation of big tech • Advocate for UBI
35. “Good” Universal Basic Income • Not just any Universal Basic Income -- Good UBI ◦ “Bad UBI” could be ▪ Merely sustenance levels of income ▪ Discourage reskilling ▪ Discourage small business ▪ Encourage dependence on big business, more corporate power ▪ Discourage civic engagement • Bad UBI could appear via deception, slippery slopes ◦ May be difficult to reverse • “Good UBI”: ◦ Protection from technological employment ◦ Without ceding more power to corporate interests ◦ Providing opportunities for reskilling, small business ◦ Encourage civic engagement See: Manna – Two Views of Humanity’s Future – by Marshall Brain https://marshallbrain.com/manna1