1. Introduction Tamara Berg CS 590-133 Artificial Intelligence Many slides throughout the course adapted from Dan Klein, Stuart Russell, Andrew Moore, Svetlana Lazebnik, Percy Liang, Luke Zettlemoyer

2. Today Course Info What is AI? History of AI Current state of AI

3. Course Information Instructor: Tamara Berg (tlberg@cs.unc.edu)tlberg@cs.unc.edu Office Hours: FB 236, Tues/Thurs 4:45-5:45pm Course website: http://tamaraberg.com/teaching/Spring_14/http://tamaraberg.com/teaching/Spring_14/ Course mailing list: comp590-133@cs.unc.educomp590-133@cs.unc.edu TAs: Shubham Gupta & Rohit Gupta TA office hours: TBD Announcements, readings, schedule, etc, will all be posted to the course webpage. Schedule may be modified as needed over the semester. Check frequently!

4. Course Information Textbook: “Artificial Intelligence A Modern Approach” Russell & Norvig, 3 rd edition Prerequisites: – Programming knowledge and data structures (COMP 401 and 410) are required – Reasonable familiarity with probability, algorithms, calculus also highly desired – There will be a lot of math and programming Work & Grading – Readings (mostly from textbook) – 5-6 assignments including written questions, programming, or both – 2 midterms (approximate dates are on course website) and final exam – Grading will consist of 60% assignments, 40% exams. For borderline cases participation in class or via the mailing list may also be considered.

5. Programming Students are expected to know how to program. Programming assignments will be in python – useful language to know, used in many current AI courses, not too hard to pick up given previous programming experience. First week – install python on your laptops, do a python tutorial TAs will also hold drop in tutorials early next week (probably Mon/Tues, times will be posted to website). Make sure to attend if you’re new to python or want a refresher.

6. Course Information Late policy: Assignments must be turned in electronically by 11:59pm on the listed due date. Students will be allowed 5 free homework late days of their choice over the semester (you don't need to ask ahead of time, just use them and we will keep track). After those are used late homework will be accepted up to 1 week late, with a 10% reduction in value per day late.

7. Course Information Honor code: Students are encouraged to complete the assignments in groups of 2. You may discuss problems at a high level with other students in the class, but all code and written responses should be original within your pair. To protect the integrity of the course, we will actively check for code or written plagiarism (both from current classmates and the internet). Exams will be closed book.

8. About me 1997-2001 Undergrad at U.W. Madison CS and Math 2001-2007 Grad at U.C. Berkeley Ph.D. in CS 2007-2008 Postdoc at Yahoo! Research 2008-2013 Assistant Prof at SBU 2013- Assistant Prof at UNC

9. My research interests Yellow lady’s slipperleonberg Image Classification 1000 classes 62.5% accuracy, Krizhevsky et al Human-centric Computer Vision Object Detection 20 object classes 39% accuracy, Girshick et al Image Parsing 33 labels 55% accuracy, Tighe et al Computer Vision

10. This is a picture of one sky, one road and one sheep. The gray sky is over the gray road. The gray sheep is by the gray road. Here we see one road, one sky and one bicycle. The road is near the blue sky, and near the colorful bicycle. The colorful bicycle is within the blue sky. BabyTalk: Generating natural language image descriptions This is a picture of two dogs. The first dog is near the second furry dog.

11. Recognizing Clothing

12. Application: Pose Independent Retrieval

13. About you? Undergrad/grad Year Major/Minors Background in: – Programming – Calculus – Probability – Python

14. Sci-Fi AI

15. AI Knowledge representation Planning Learning Natural Language Processing Perception (computer vision, speech) Reasoning Motion & Manipulation (robotics) Social Intelligence Creativity

16. What is AI? Definitions of AI: 1. Thinking humanly2. Acting humanly 3. Thinking Rationally4. Acting rationally

17. AI definition 1: Thinking humanly Need to study the brain as an information processing machine: cognitive science and neuroscience

18. AI definition 1: Thinking humanly Can we build a brain?

19. AI definition 1: Thinking humanly Can we build a brain?

20. The Turing Test What capabilities would a computer need to have to pass the Turing Test? – Natural language processing – Knowledge representation – Automated reasoning – Machine learning AI definition 2: Acting humanly A. Turing, Computing machinery and intelligence, Mind 59, pp. 433-460, 1950Computing machinery and intelligence

21. Turing predicted that by the year 2000, machines would be able to fool 30% of human judges for five minutes Loebner prize – 2008 competition: each of 12 judges was given five minutes to conduct simultaneous, split-screen conversations with two hidden entities (human and chatterbot). The winner, Elbot of Artificial Solutions, managed to fool three of the judges into believing it was human [Wikipedia].Wikipedia The Turing Test

22. Success depends on deception! Chatbots can do well using “cheap tricks” First example: ELIZA (1966)ELIZA Chinese room argument: one may simulate intelligence without having true intelligence (more of a philosophical objection) Chinese room argument Turing Test: Criticism

23. A better Turing test? http://www.newyorker.com/online/blogs/elements/2013/08/why-cant-my- computer-understand-me.html

24. A better Turing test? Multiple choice questions that can be easily answered by people but cannot be answered by computers using “cheap tricks”: The trophy would not fit in the brown suitcase because it was so small. What was so small? The trophy The brown suitcase H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

25. A better Turing test? Multiple choice questions that can be easily answered by people but cannot be answered by computers using “cheap tricks”: The trophy would not fit in the brown suitcase because it was so large. What was so large? The trophy The brown suitcase H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

26. A better Turing test? Multiple choice questions that can be easily answered by people but cannot be answered by computers using “cheap tricks”: The large ball crashed right through the table because it was made of styrofoam. What was made of styrofoam? The large ball The table H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

27. A better Turing test? Multiple choice questions that can be easily answered by people but cannot be answered by computers using “cheap tricks”: The large ball crashed right through the table because it was made of steel. What was made of steel? The large ball The table H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

28. A better Turing test? Multiple choice questions that can be easily answered by people but cannot be answered by computers using “cheap tricks”: Sam tried to paint a picture of shepherds with sheep, but they ended up looking like golfers. What looked like golfers? The shepherds The sheep H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

29. A better Turing test? Multiple choice questions that can be easily answered by people but cannot be answered by computers using “cheap tricks”: Sam tried to paint a picture of shepherds with sheep, but they ended up looking like rabbits. What looked like rabbits? The shepherds The sheep H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

30. A better Turing test? Why are these questions hard for computers?? H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

31. A better Turing test? Advantages over standard Turing test Test can be administered and graded by machine Does not depend on human subjectivity Does not require ability to generate English sentences Questions cannot be evaded using verbal dodges Questions can be made “Google-proof” H. Levesque, On our best behaviour, IJCAI 2013On our best behaviour

32. AI definition 3&4: Rationality A rational agent acts to optimally achieve its goals Goals are application-dependent and are expressed in terms of the utility of outcomes Being rational means maximizing your (expected) utility This definition of rationality only concerns the decisions/actions that are made, not the cognitive process behind them In practice, utility optimization is subject to the agent’s computational constraints (bounded rationality or bounded optimality)

33. Utility maximization formulation Advantages – Generality: goes beyond explicit reasoning, and even human cognition altogether – Practicality: can be adapted to many real-world problems – Naturally accommodates uncertainty – Amenable to good scientific and engineering methodology – Avoids philosophy and psychology Disadvantages?

34. History of AI Image source

35. Origins of AI: Early excitement 1940s First model of a neuron (W. S. McCulloch & W. Pitts) Hebbian learning rule Cybernetics 1950sTuring Test Perceptrons (F. Rosenblatt) Computer chess and checkers (C. Shannon, A. Samuel) Machine translation (Georgetown-IBM experiment) Theorem provers (A. Newell and H. Simon, H. Gelernter and N. Rochester) 1956Dartmouth meeting: “Artificial Intelligence” adopted

36. Herbert Simon, 1957 “It is not my aim to surprise or shock you – but … there are now in the world machines that think, that learn and that create. Moreover, their ability to do these things is going to increase rapidly until – in a visible future – the range of problems they can handle will be coextensive with the range to which human mind has been applied. More precisely: within 10 years a computer would be chess champion, and an important new mathematical theorem would be proved by a computer.” Simon’s prediction came true – but forty years later instead of ten

37. Harder than originally thought 1966: Eliza chatbot (Weizenbaum)Eliza “ … mother …” → “Tell me more about your family” 1954: Georgetown-IBM experimentGeorgetown-IBM experiment Completely automatic translation of more than sixty Russian sentences into English Only six grammar rules, 250 vocabulary words, restricted to organic chemistry Promised that machine translation would be solved in three to five years (press release)press release Automatic Language Processing Advisory Committee (ALPAC) report (1966): machine translation has failed “The spirit is willing but the flesh is weak.” → “The vodka is strong but the meat is rotten.”

38. Blocks world (1960s – 1970s) Larry RobertsLarry Roberts, MIT, 1963 ???

39. History of AI: Taste of failure 1940s First model of a neuron (W. S. McCulloch & W. Pitts) Hebbian learning rule Cybernetics 1950s Turing Test Perceptrons (F. Rosenblatt) Computer chess and checkers (C. Shannon, A. Samuel) Machine translation (Georgetown-IBM experiment) Theorem provers (A. Newell and H. Simon, H. Gelernter and N. Rochester) Late 1960sMachine translation deemed a failure Neural nets deprecated (M. Minsky and S. Papert, 1969)* Late 1970sThe first “AI Winter”“AI Winter” * A sociological study of the official history of the perceptrons controversy A sociological study of the official history of the perceptrons controversy

40. History of AI to the present day 1980sExpert systems boom Late 1980s- Expert system bust; the second “AI winter” Early 1990s Mid-1980s Neural networks and back-propagation Late 1980sProbabilistic reasoning on the ascent 1990s-PresentMachine learning everywhere Big Data Deep Learning Building Smarter Machines: NY Times Timeline AAAI Timeline History of AI on Wikipedia

41. NY Times article

42. What accounts for recent successes in AI? Faster computers The IBM 704 vacuum tube machine that played chess in 1958 could do about 50,000 calculations per second Deep Blue could do 50 billion calculations per second – a million times faster! Dominance of statistical approaches, machine learning Big data Crowdsourcing

43. Historical themes Moravec’s paradox “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility” [Hans Moravec, 1988] Why is this? Early AI researchers concentrated on the tasks that they themselves found the most challenging, abilities of animals and two-year-olds were overlooked We are least conscious of what our brain does best Sensorimotor skills took millions of years to evolve, whereas abstract thinking is a relatively recent development

44. Historical themes Silver bulletism (Levesque, 2013):Levesque, 2013 “The tendency to believe in a silver bullet for AI, coupled with the belief that previous beliefs about silver bullets were hopelessly naïve” Conceptual dichotomies (Newell, 1983):Newell, 1983 Symbolic vs. continuous High-level vs. low-level modeling of mental processes Serial vs. parallel Problem solving vs. recognition Performance vs. learning Boom and bust cycles Periods of (unjustified) optimism followed by periods of disillusionment and reduced funding Image problems AI effect: As soon as a machine gets good at performing some task, the task is no longer considered to require much intelligence AI effect

45. What can AI do today?

46. IBM Watson http://www-03.ibm.com/innovation/us/watson/ NY Times article Trivia demo IBM Watson wins on Jeopardy (February 2011)IBM Watson wins on Jeopardy

47. Self-driving cars Google’s self-driving car passes 300,000 miles (Forbes, 8/15/2012) Google’s self-driving car passes 300,000 miles Nissan pledges affordable self-driving car models by 2020 (CNET, 8/27/2013) Nissan pledges affordable self-driving car models by 2020

48. Natural Language Speech technologies Google voice search Apple Siri Machine translation translate.google.com Comparison of several translation systems

49. Vision OCR, handwriting recognition Face detection/recognition: many consumer cameras, Apple iPhotoApple iPhoto Visual search: Google Goggles, search by imageGoogle Gogglessearch by image Vehicle safety systems: MobileyeMobileye

50. Mathematics In 1996, a computer program written by researchers at Argonne National Laboratory proved a mathematical conjecture unsolved for decades NY Times story: “[The proof] would have been called creative if a human had thought of it” NY Times story Mathematical software:

51. Games IBM’s Deep Blue defeated the reigning world chess champion Garry Kasparov in 1997 1996: Kasparov Beats Deep Blue “I could feel – I could smell – a new kind of intelligence across the table.” 1997: Deep Blue Beats Kasparov “Deep Blue hasn't proven anything.” In 2007, checkers was “solved” (though checkers programs had been beating the best human players for at least a decade before then) Science article

52. Logistics, scheduling, planning During the 1991 Gulf War, US forces deployed an AI logistics planning and scheduling program that involved up to 50,000 vehicles, cargo, and people NASA’s Remote Agent software operated the Deep Space 1 spacecraft during two experiments in May 1999Remote Agent In 2004, NASA introduced the MAPGEN system to plan the daily operations for the Mars Exploration RoversMAPGEN

53. Information agents Search engines Recommendation systems Spam filtering Automated helpdesks Fraud detection Automated trading Medical diagnosis

54. Robotics Mars rovers Autonomous vehicles – DARPA Grand Challenge DARPA Grand Challenge – Self-driving cars Autonomous helicopters Robot soccer – RoboCup RoboCup Personal robotics – Humanoid robots Humanoid robots – Robotic pets Robotic pets – Personal assistants?

55. Towel-folding robot J. Maitin-Shepard, M. Cusumano-Towner, J. Lei and P. Abbeel, Cloth Grasp Point Detection based on Multiple-View Geometric Cues with Application to Robotic Towel Folding, ICRA 2010Cloth Grasp Point Detection based on Multiple-View Geometric Cues with Application to Robotic Towel Folding More clothes folding YouTube Video

56. Outline of the Class Part 1: Agents & Decisions – Fast search – Constraint satisfaction – Reinforcement Learning Part 2: Modeling Uncertainty – Probability – Bayes Nets Part 3: Learning from labeled Data – Classification Part 4: Sub-Areas of AI – NLP – Vision

57. Philosophy of the class Our goal is to use machines to solve hard problems that traditionally would have been thought to require human intelligence We will try to follow a sound scientific/engineering methodology Consider relatively limited application domains Use well-defined input/output specifications Zero in on essential problem features Focus on principles and basic building blocks

58. For next week Check out the class website http://www.tamaraberg.com/teaching/Spring_14/ http://www.tamaraberg.com/teaching/Spring_14/ Get the book. Do the readings. Do a python tutorial. TAs will hold an in person drop-in tutorial. Dates/times will be posted to the class website (probably Mon&Tues evening).