COMP-424: Artificial intelligence. Lecture 2: Introduction to AI!

Transcription

1 COMP Artificial Intelligence Lecture 2: Introduction to AI! Instructor: Joelle Pineau (jpineau@cs.mcgill.ca) Class web page: Unless otherwise noted, all material posted for this course are copyright of the instructor, and cannot be reused or reposted without the instructor s written permission. 2 1

2 Outline for today Recap of course basics Overview of AI history Approach to teaching AI Examples of AI applications 3 Artificial Intelligence Planning Reasoning Decisionmaking Search Learning Perception Modelling Language Knowledge 4 2

3 About myself Education: BASc in Engineering (U.Waterloo) MSc / PhD in Robotics (Carnegie Mellon U.) Prof at McGill What kind of AI research do I do? Machine learning Planning and decision-making Robotics Personalized medicine and health-care 5 Research areas in my lab Reinforcement learning Supervised learning Representation learning Education MDP / POMDP Algorithms Other applications Planning Sequential Decision- Making Problems Healthcare Robotics Dynamic treatment regimes Adaptive trials Event prediction Smart wheelchairs Social navigation Marketing Resource management Industrial processes Human-robot interaction 6 3

4 Research focus Main scientific goal: Design AI agents (robots) that exhibit intelligent behavior by providing them with the ability to learn and reason. Abilities Goals/Preferences Prior Knowledge AI agent Observations Actions Environment Main tools: Probability theory, statistics, optimization, analysis of algorithms, numerical approximations, robotics, 7 SmartWheeler project COMP-102: Computers and Computing 8 4

5 About You 197 registered, from a variety of departments / faculties Computer Science Software / Electrical / Mechanical / Chemical Engineering Mathematics / Statistics / Physics Biology / Physiology / Anatomy / Microbiology & Immunology Neuroscience / Cognitive science Linguistics / Philosophy Economics / Finance Music Undeclared / Exchange 9 About the teaching team (subject to change) Office hours every day! Eric Crawford: Mon. 2-3pm MC 108 Joelle Pineau: Tues am MC 106N Harsh Satija: Wed am MC 106 Stephanie Laflamme: Thurs. 1-2pm MC 112 Emmanuel Bengio: Fri. 11am-12pm MC 111 Juan Camilo Gamboa Higuera: Fri. 2-3pm MC 400 See mycourses discussion board for any changes to days/times. (Other times possible by appointment.) Best way to contact us: (see course syllabus for addresses) Don t text, don t call. Have reasonable expectations about response time (24 hours). 10 5

6 About the course: Comments from 2014 Way more boring than I had expected. My interest in machine learning and artificial intelligence is significantly less as a result of the course. I do not recommend this course for people who intend to make use of artificial intelligence or machine learning in the future as one may develop a distaste for the material. Needs more prerequisites. The entire second half of the course is extremely difficult to follow without prior experience with statistics and probability. Too much focus on probability instead of actual AI. The workload for the course (which is 3 credits) was beyond 4-credit level. This course would be better if it did not attempt to appeal to a large class size. This is a 400 level course and cannot be taught like a 200 level intro course. The class was far too large. The TAs were so tight on time that they could barely provide any feedback on the assignments. The only feedback I've received was basically "see the solutions. 11 About the course: Comments from 2015 This course demoralized me and almost made me want to switch my major to philosophy just to get out of the complicated hell that is AI. But it is pretty cool, just extremely difficult. The warning was given, but the post-midterm material was quite disappointing and made it really hard to concentrate. This is the direction that AI has taken though, so it is hard to complain. Interesting course, but a lot of work needs to be put in for only 3 credits. The content was explained to be dry and tedious in the first class, and it was indeed dry and tedious. 12 6

7 Read this carefully Some of the course work will be individual, other components can be completed in groups. It is the responsibility of each student to understand the policy for each work, and ask questions of the instructor if this is not clear. It is the responsibility of each student to carefully acknowledge all sources (papers, code, books, websites, individual communications) using appropriate referencing style when submitting work. We will use automated systems to detect possible cases of text or software plagiarism. Cases that warrant further investigation will be referred to the university disciplinary officers. Students who have concerns about how to properly use and acknowledge third-party software should consult the course instructor or TAs. COMP-598: Applied Machine Learning 13 On the use of mobile devices in class Many bring mobile computing and communication devices to class. This can be useful! Taking notes, supplementing course material. This can be distracting! For you, for me, for your colleagues. Policy for this course: Use is encouraged during quizzes. Use is tolerated during lectures in support of learning. Use for other purposes happens outside of the classroom. For more discussion of this issue, see: Commun-devices-MC2-guidelines-11June2010.pdf 14 7

8 Questions? COMP-598: Applied Machine Learning 15 What is biological intelligence? Sensory processing: Visual cortex. Auditory cortex. Somatosensory cortex. Motor cortex. Cognitive functions: Memory. Reasoning. Executive control. Learning. Language. 16 8

9 What is biological intelligence? A mix of general-purpose and special-purpose algorithms. General-purpose: Memory formation, updating, retrieval. Learning new tasks. Special-purpose: Recognizing visual patterns. Recognizing sounds. Learning language. All are integrated seamlessly! 17 What is AI? 18 9

10 What is AI? Some answers: Modeling human cognition using computers. Studying problems that others don t know how to solve. Cool stuff! Game playing, machine learning, data mining, speech recognition, computer vision, web agents, robots. Useful stuff! Medical diagnosis, fraud detection, genome analysis, object identification, space shuttle scheduling, information retrieval. Solving AI is potentially the meta-solution to all problems! 19 What is AI? My (current) best answer: Developing models and algorithms that can produce rational behaviors in response to incoming stimulus and information

11 What is AI? Human intelligence: Artificial Intelligence: Sensory processing: Visual cortex.!!computer vision Auditory cortex.!!signal/speech processing Somatosensory cortex.!!haptics Motor cortex.!!robotics Cognitive functions Memory.!Knowledge representation. Reasoning.!!Search, inference. Executive control.!!planning, decision-making. Learning.!Model learning. Language.!Language understanding. 21 Different goals of AI This would be pretty awesome! Thinking Humanly Acting Humanly Thinking Rationally Acting Rationally 22 11

12 Acting humanly AI is about duplicating what the (human) brain DOES. Alan Turing ( ) had interesting thoughts about this. Can a machine think? -> If it could, how would we tell? Turing (1950): Computing machinery and intelligence Human interrogator? Human AI agent An operator interacts with either the human or the AI agent. Can he correctly guess which one? 23 Turing s prediction By 2000, a machine would have a 30% chance of fooling a lay person for 5 minutes. Suggested major components of AI: Knowledge representation, automated reasoning, language understanding, machine learning But here s a thought: We succeeded in building flying machines when we stopped trying to imitate birds, and started learning about aerodynamics

13 Different goals of AI Thinking Humanly Acting Humanly Thinking Rationally Acting Rationally Let s try this! 25 Acting rationally Rational behaviour = doing the right thing. Doing what is expected to maximize goal achievement, given the available information and available resources. Does not necessarily require thinking (e.g. blinking reflex). But in many cases, thinking serves rational behaviour. This is the flavour of AI we will focus on

14 Big picture Action THE WORLD Task & Environment AI Agent Perception 27 Rational agents This course is about designing rational agents. An agent is an entity that perceives and acts. Goal: Learn a function mapping percept histories to actions: f : P h A A rational agent implements this function such as to maximize performance. Performance measures: goal achievement, resource consumption,... Caveat: Many constraints can come into play (time, space, energy, bandwidth, ) which make perfect rationality unachievable. Objective: Find best function for given information and resources

15 AI Beginnings ENIAC: First super-computer, created in Early work in 1950s: Rosenblatt s perceptron Samuel s checkers player 29 Dartmouth Conference (1956) 30 15

16 Dartmouth Conference (1956) Some of the attendees: John McCarthy: LISP, time-sharing, application of logic to reasoning Marvin Minsky: popularized neural networks and showed their limits, introduced slots and frames Claude Shannon: information theory, juggling machine Allen Newell and Herb Simon: bounded rationality, general problem solver, SOAR The meeting coined the term artificial intelligence'' 31 Early AI hopes and dreams Make programs that exhibit similar signs of intelligence as people: prove theorems, play chess, have a conversation. Logical reasoning was key. Learning from experience was considered important. The research agenda was geared towards building general problem solvers. There was a lot of hope that natural language could be easily understood and processed

17 Recent AI: Math to the rescue! Heavy use of probability theory, decision theory, statistics. Trying to solve specific problems rather than aim for general reasoning. AI today is a collection of sub-fields: Perception and computer vision. Natural language understanding. Robotics Etc. Reasoning is now the part named ``AI''. A lot of progress was made in this way! Some recent efforts try to put all this together (e.g. in robotics). 33 Example AI system (1997): Chess playing IBM Deep Blue defeats Garry Kasparov. Perception: advanced features of the board. Actions: choose a move. Reasoning: search and evaluation of possible board positions

18 Example AI system (2008): Poker playing University of Alberta s Polaris defeats some of the world s best online pros. One variety of poker: Heads-up limit Texas Hold em (two players, limited betting amounts) Perception: features of the game. Actions: choose a move. Reasoning: search and evaluation of possible moves, machine learning Example AI system (2011): Jeopardy! 36 18

19 Example AI system (2015): Atari 37 AI meets markets Trading agent competition ( ): Trades Stock market Trading Agent Rates, news 38 19

20 AI meets markets 39 Example AI system (1992): Medical diagnosis Pathfinder (D. Heckerman, Microsoft Research) Perception: symptoms, test results. Actions: suggest tests, make diagnosis. Reasoning: bayesian inference, machine learning, Monte-Carlo simulation

21 Example AI system (2008): Reading the Mind Brain Image Analysis (T. Mitchell, CMU) Perception: brain imaging using fmri technology. Actions: detect which word (e.g. hammer, apartment, ) is being read by the human subject. Reasoning: statistical machine learning Example AI system (2003): Digit recognition Handwritten digit recognition: <1% error. Previously seen known examples New example to classify... Boxes represent the weights into a hidden node in a neural network learner. COMP-424: COMP-598: Artificial Applied Machine intelligence Learning 42 21

22 Example AI system (2015): Object recognition ImageNet Large Scale Visual Recognition Challenge Image database with 1000 object categories, hundreds of example images for each class, over 1 million images in total. Current best results: ~5% classification error (top 5 categories). 43 Example AI system (1998): Automatic driver ALVINN (D. Pomerleau, CMU) drives autonomously for 21 miles on the highway at speeds of up to 55 miles/hour. Perception: digitized camera images of the road. Actions: 64 steering angles. Reasoning: artificial neural network

23 Example AI system (2005): Off-road driving Stanley (Stanford University) navigates 132 miles through desert terrain in less than 10 hours, with no human intervention. Perception: GPS, 6-D inertial measurement, wheel speed, 4 lasers, 1 radar, stereo and monocular cameras. Actions: actuation of drive-by-wire system. Reasoning: position estimation, path planning Example AI system (2014): City driving Google cars have logged over 700,000 miles in autonomous mode without any accident. Sensors and actuators similar to Stanley (GPS, 3D laser point cloud, cameras, odometry) Significant prior knowledge: City of Mountain View (12-square-mile) is fully mapped at high resolution

24 AI and the Web Google: 3.5 billion searches per day. Facebook: 2 billion photos uploaded every day. 700 engineers building machine learning algorithms to analyze the data. YouTube: 300 hours of video uploaded every minute. Baidu: 500 million page views per day Flickr: 2 million photos uploaded every day. Twitter: 500 million tweets sent per day. 47 The Future of AI Quotes from this week s symposium: IBM Senior VP John Kelly: Nothing I have ever seen matches the potential of AI and cognitive computing to change the world. NVIDIA founder Jen-Hsun Huang: AI is the final frontier [..] I ve watched it hyped so many times, and yet, this time, it looks very, very different to me. Former Google CEO and now Chairman of Alphabet Eric Schmidt: 1. AI should benefit the many, not the few. 2. AI R&D should be open, responsible and socially engaged. 3. Developers of AI should establish best practices to minimize risks and maximize the beneficial impact

25 Final comments Come to class! Come prepared! Be willing to learn and work hard! For next class: Read chapters 1-3 in Russell & Norvig. Review the course syllabus. Questions? 49 25