Utopia/dystopia: The good, the bad and the basics of machine learning

A general introduction to AI and machine learning. Describes the difference between ML and and traditional algorithms. Gives and introduction to some of the basic building blocks of ML models. Provides some examples of successful applications. Considers some possible dangers of artificial general intelligence and artificial vertical intelligence.

Transcript

1. Utopia/Dystopia The good, the bad and the basics of Machine Learning November 2017 Keith T. Butler
2. Our robot overlords
3. • What is machine learning? • What are the applications of machine learning? • How can I do machine learning? • What are the dangers of machine learning? Outline
4. What is machine learning Animals learn from experience Machines are programmed Machines learn from experience data
5. Procedural logic – a list of instructions on how to do things. An algorithm
6. Deep blue Evaluation function
7. Deep blue Minimax algorithm
8. Alpha Go
9. • A type of AI • Learning, without explicit programming • Learning on data Machine learning
10. Why is ML is big right now?
11. Machine learning hype www.gartner.com
12. Some examples of ML success
13. Some examples of ML success http://intelligentx.ai/
14. AI and craft beer http://aiweirdness.com/post/163753995072/craft-beer-names-invented-by-neural-network IPA Amber Stout Dang River Frog Trail Ale Cherry Trout Stout Juicy Dripple IPA O'Busty Irish Red The Bopberry Stout Earth Pump Ricias Donkey Brain Barrel Aged Chocolate Milksmoke
15. Domain experts needed!
16. Democratic Technology Open education + Open data + Open software = Democratic technology
17. Open education Straight into making models Some programming required High-school maths Builds from the bottom Teaches Python and data science Coursera – provides Machine Learning course from Stanford More academic and rigorous
18. Open software
19. Open data
20. A good start
21. The goal of machine learning To generalise beyond the training set. Doing well on the training set is easy.
22. Learning = Representation + Evaluation + Optimisation Learning Representation How we represent the knowledge. This also chooses the set of possible classifiers. Hypothesis space. Eg. Neural network, decision tree …
23. Learning = Representation + Evaluation + Optimisation Learning Evaluation Objective function or scoring function. Distinguish good from bad classifiers. NB need not be the same as the external function that the classifier is optimising.
24. Learning = Representation + Evaluation + Optimisation Learning Optimisation Searches between classifiers. Identifies the highest-scoring one. Determines the efficiency of a learner.
25. Learning table "A Few Useful Things to Know About Machine Learning" by Pedro Domingos
26. Linear regression £20,000 £120,000 ? https://www.autodraw.com/
27. Linear regression 20k 120k 70k
28. Naïve Bayes
29. Naïve Bayes Spam Not spam
30. Naïve Bayes Cheap Spam Not spam
31. Naïve Bayes Cheap Spam Not spam 80% chance that an email with the word ‘cheap’ will be spam
32. Naïve Bayes Cheap Typos Caps title 80 % 60 % 90 % 99.2 % CHEAP PREGNENCY TEST
33. Decision tree
34. Decision tree Age Beard Choice 56 Y 49 N 25 Y 32 N 38 Y 63 Y
35. Decision tree Age Beard Choice 56 Y 49 N 25 Y 32 N 38 Y 63 Y Which is a more useful descriptor?
36. Decision tree Age Beard Choice 56 Y 49 N 25 Y 32 N 38 Y 63 Y Which is a more useful descriptor?
37. Decision tree Age Beard Choice 56 Y 49 N 25 Y 32 N 38 Y 63 Y Age > 45 < 45 Beard? Y N
38. Decision tree Age > 45 < 45 Beard? Y N
39. Neural network Hops Malt Yes No
40. Neural network
41. Neural network
42. Problem 1: Induction Data alone is not enough – we need knowledge.
43. A consequence of induction No learner can beat random guessing over all possible functions to be learned.
44. Getting a free breakfast! Assumptions Eg. Smoothness, similar examples have similar classes etc. Representation choice Eg. If we have a lot of knowledge about the preconditions required by a class, then ‘IF…THEN’ representation may be best.
45. Problem 2: Fitting Bias Constantly learning the same wrong thing. Variance Learn random things irrespective of real values.
46. An example Voting patterns http://scott.fortmann-roe.com/docs/BiasVariance.html
47. An example k-nearest neighbours http://scott.fortmann-roe.com/docs/BiasVariance.html
48. An example k = 83 http://scott.fortmann-roe.com/docs/BiasVariance.html
49. An example k = 1 http://scott.fortmann-roe.com/docs/BiasVariance.html
50. An example k = 12 http://scott.fortmann-roe.com/docs/BiasVariance.html
51. Overfitting bed
52. Cross validation N-fold cross validation
53. • Raw data -> something meaningful • The vast majority of ML time is spent here Problem 3: representing features
54. AI branches Artificial intelligence Artificial general intelligence Artificial vertical intelligence
55. AGI and the problem of prediction I don't have an answer for the obstacles in our way to building human-like robots, because we just don't know enough about how people reason. We haven't figured out the fundamental principles, so I don't think we know what the hurdles are. “But the real future of the laptop computer will remain in the specialized niche markets. Because no matter how inexpensive the machines become, and no matter how sophisticated their software, I still can't imagine the average user taking one along when going fishing.” NY Times 1985 “That the automobile has practically reached the limit of its development is suggested by the fact that during the past year no improvements of a radical nature have been introduced.” Scientific American 1909
56. ML: decision making and agency
57. Tackling/understanding bias https://research.google.com/bigpicture/attacking-discrimination-in-ml/
58. Parole software https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
59. Public scrutiny and awareness https://ainowinstitute.org/ https://algorithmwatch.org/en/ https://www.propublica.org/
60. Utopia/Dystopia? The Goliath of totalitarianism will be brought down by the David of the microchip. Ronald Regan 1989
61. http://people.bath.ac.uk/ktb22/ Thank You https://github.com/keeeto https://www.mixcloud.com/keith-butler2/ @keeeto2000