2 After this module, you will be able to… Explain the details of the Turing TestExplain counter arguments to the Turing TestList and describe several AI techniquesUnderstand the implications of AI
3 Can computers be intelligent? What is “intelligence”?from the dictionary: “The ability to learn from experience, to reason, and to think.”Can computers do this, or be made to??and if they can, does that make them intelligent??
4 Intelligence in a computer The basic metaphormind = softwarebrain = computerMind vs. softwarehuman minds receive input, generate output, and process and remember informationso does softwaretaking this metaphor seriously means building software to emulate minds
5 Brain versus computer Processing power and speed can only estimate the processing power of the average human brain (at this time)best guess: nerve volume to be proportional to processing powerretina processes about ten one-million-point images per second1,500cc human brain is about 100,000 times as large as the retina, therefore average brain about 100 million MIPS (Million computer Instructions Per Second ) while recent super-computers have only a few million MIPS worth of processor speed
6 Brain versus computer Memory (have you ever ‘run out’ of memory?) best guesscalculating using the number of synapses connecting the neurons in the human brain and estimate each of them to be capable holding one megabyte worth of memory. Since the brain has 100-trillion-synapses, we can estimate that the average brain can hold about 100 million megabytes of memory (100 terabytes)Versus: computer: 2 GB RAM, 300 GB Harddrive1 terabyte = 1024 Gigabytes (GB)Terabyte Harddrive: $600 US
7 Brain versus computer Performance and ability a matter of programming?The brain is made for general purposes, not specifically just for computational jobsin theory, the brain could be as quick as a computer in computational and recording jobs but the average human is constantly distracted by their senses, emotions & thoughts
8 Brain versus computer Adaptability and learning Time to learn the human brain is able to learn by trial and error, induce conclusions from past experience and create new methods to deal with the situations - highly adaptable!Time to learnhowever, the human brain requires time to adapt to the behaviors required when learning a new skill or quitting a old habitThe computer is getting better every year - don’t forget Moore’s Law!
9 Defining Artificial Intelligence Artificial intelligence is the study of ideas which enable computers to do the things that make people seem intelligent.—Patrick Henry Winston, in Artificial IntelligenceArtificial intelligence is the study of how to make computersdo things at which, for the moment, people are better.Artificial intelligence is the study of the computationsthat make it possible to perceive, reason, and act.—Patrick Henry Winston, in Artificial IntelligenceArtificial intelligence is the study of those things we don’t yet know how to make the computer do.
10 a computer is intelligent? How do we measure ifa computer is intelligent?A machine may be deemed intelligent when it can pass for a human being in a blind test.—Alan TuringThe Turing Test
11 Alan Turing’s “Turing Test” Turing proposes thatIf a computer acts intelligently then it is intelligent.If it walks like a duck, quacks like a duck and flies like a duck, it has to be a duckRemoves all physical characteristics from considerationBased entirely on “intelligence”Never been passed yetTuring thought it would be passed by 2000 but it hasn’tThe Loebner contest: $100K prize for 1st program to pass the Turing test (no winners so far!)
12 How the Turing Test works Invented in the 1950’sPurpose: to test if a computer is “intelligent”Set-up:requires 2 humans and a computerone human is the “Interrogator” and kept in a separate roomThe Interrogator has a typed conversation with both the human and the computer, without knowing which is whichThe computer is considered intelligent if the interrogator fails to distinguish between the responses of the human and those of the computer
13 Falsifying the turing test Searle’s Chinese Roomgiant room in China with input and output slotsmessages written in Chinese are put into input slotanswer, also in Chinese, arrives in output slotanswer cannot be distinguished from answers a (Chinese speaking) human would provideWhat’s inside the room?if there is a human there, then we would say “the room” exhibits intelligenceif there is a computer there, we would say it has passed the Turing Test
14 Falsifying the turing test But, what’s inside the room isan army of unilingual English clerks and a huge number of filing cabinets filled with cards with various rules on them mapping one Chinese symbol into other Chinese symbols (it’s a BIG room!)When a question arrives in the input slotit is broken down into its constituent symbols by the clerks, each symbol written on a piece of paperfor each symbol, clerks look up the cards associated with the symbol, copying new symbols onto new pieces of paper based on the rules on the cardsthese new symbols lead to the look up of further cards and the application of rules on these cards creating still more pieces of paper with more symbolsthis process iterates, mediated by the clerks, until it terminates with a final piece of paper that is pushed through the output slot
15 Falsifying the turing test This is an exact analogy of the Turing TestThe process inside the room is an exact analogy for how a computer processes informationthe clerks are like processing unitsthe filing cabinets are like computer memorythe cards with the rules are like programsSo, where is the intelligence?in the clerks? - but they are just following orders without intelligence (besides, they don’t speak Chinese)in the cabinets or the cards or the room? - but they are inanimatewhere else?Must, therefore, mean that the Turing Test is wrongOr does it?
16 Making an intelligent computer Two common approaches to AI:one approach attempts to use computers to simulate human mental processes.we don’t really know how brains work – hard to simulate!brain much more complex than most powerful computertwo main typescognitive psychology (simulating mind)neural networks (simulating brain)the second, more common, approach to AI involves designing intelligent machines, independent of the way people think.acknowledge that computers are adept at solving problems in other ways
17 Game playingOne of the most successful areas of AI has been game playingChess: Deep Blue has beaten Gary Kasparov, the reigning world championCheckers: the world’s best checkers player is a computer program, Chinook (built at U. of Alberta)Poker: now being exploredGo: still too complexTechniques work best for complete information games with no element of chanceRock Paper Scissors:
18 Lookahead search Main game playing technique is lookahead search extrapolate all possible moves from the current board position “growing” a game treethen examine the bottom of the tree for (extrapolated) board positions that look good (using an evaluation function)make a move that would lead towards the best of these positionsrepeat this process after the opponent has movedestimated game tree sizes, if generated to the various end games:Tic-Tac-Toe: 9! = 362,880 nodesCheckers: 10**78 nodesChess: 10**120 nodesGo: 10**720 nodesestimated duration of the universe to date (in nano-seconds):4.5*10**27 (this is less than 10**28)game playing researchers have found lots of techniques for “pruning” game trees to make lookahead search more tractableHumans use strategies that do relatively little lookahead
20 Searching in other domains AI techniques developed by game playing researchers have led to general searching techniques in a variety of applications outside of game playing, eg.planningautomated theorem provingconstraint satisfactionAs in game playing such searches often combinatorially explode, and there thus needs to be some way to prune and guide searchesheuristics: rules of thumban evaluation function in game playing is a heuristic: allows evaluation of a game position without further growing the search treeThere is much mathematical work on proving the efficiency (computational complexity) of search algorithms, looking for better search algorithms
21 Natural Language Communication Harder than it seemsdetermining grammatical correctness (syntax)rules are vague, ambiguous, and contradictory.Colorless green ideas sleep furiouslysemantic analysisThe dog lies on his blanketThe boy lies on his testdetermining meaning from dialog context (pragmatics)The chicken is ready to eatthe problem of ambiguityThey are flying planesI saw Mt. Robson flying to VancouverTime flies like an arrow
22 Language translationThe Russian-English project (late 50’s, early 60’s)millions of dollars, eventually abandonedout of sight, out of mind – invisible idiot or blind and insanethe spirit is willing, but the flesh is weak - the vodka is good, but the meat is rottenfirst rate translation without understanding is impossibleBut, there are situations where there can be limited successestranslating scientific documents: unilingual Japanese physicist can fix up a machine-translated English language physics paperweather forecastingBabelfish:is translating legal documents such a situation?Saskatchewan’s own Gigatext disaster in the late 80’s: millions spent on translation software to translate Saskatchewan laws into French automatically - why might this not be such a good idea?
23 Conversation without communication Eliza – one of the first programs to converse in a limited form of natural language (developed by Joseph Weizenbaum in the mid-60’s)early language-processingused scripts and generic, open-ended, or neutral responses
24 Eliza goes to work.Dialogue between Eliza (upper case letters) and a simulation of a paranoid named Parry (lower case letters), developed by Kenneth Colby in the late 60’sI don't understand your motives.WHY DON'T YOU UNDERSTAND MY MOTIVES?Shouldn't I be?WHY DID YOU SAY THAT JUST NOW?That's just the way it is.SUPPOSE THAT'S JUST THE WAY IT WERE NOT.So?I'M NOT SURE I UNDERSTAND YOU FULLY.You harp on the same questions too much.WHAT DOES THAT SIMILARITY SUGGEST TO YOU?I don't get you.DO YOU WISH THAT YOU DID GET I?
25 Pattern recognitionPattern recognition involves identifying recurring patterns in input data with the goal of understanding or categorizing that inputApplications includeimage analysisoptical character recognitionautomatic speech recognitiontalking computersneural networks
26 Image analysis and vision Identifying objects and shapes in a photo, drawing, video, or other visual imageUses: colorizing classic motion pictures, piloting cruise missiles, identifying people, etc.Difficult to doirrelevant surrounding data, obstructions, indistinct edges, changes in light or angle, movement, etc.
27 Optical Character recognition Locates and identifies printed characters embedded in images (it “reads” text)About 99% accurate “at best”much better with typed text than handwritingUses: digitizing paper documents, “Palm Pilots”, translating foreign text characters…
28 Speech recognitionHave to be able to adapt to different pronunciations, accents, rate of speech, tone, voice inflection, etc.Have to be able to isolate words from a continuous signalMust be able to interpret and use all of the subtleties of speech in order to interact believablyMuch easier for the computer to speak believably than to understand speechadvancing quite fast, now possible to imitate a real voice… ethic/legal issues – can a network use an actor’s voice to say things the actor never said?
29 Speech recognitionAutomatic speech recognition systems use pattern recognition techniques similar to those used by vision and OCR systems, including these:segmentation of input sound patterns into individual words and phonemesexpert rules for interpreting soundscontext “experts” for dealing with ambiguous soundslearning from a human trainer
30 Expert reasoning systems Expert systems are programs designed to replicate the decision-making process of humanshas a knowledge base representing content from a specific fieldKnowledge bases contain facts, and a system of rules for determining and changing relationships among factscan be reorganized as new information arrives
31 Expert systemsExpert systems are difficult to build because it’s hard to capture “all” knowledge from an expertIf-then rulesif the engine will not turn over and lights don’t work, then check the batteryProbabilistic reasoningdeals with uncertainties, conclusions stated as probabilitiesthere’s a 70% chance that this patient has a bacterial infection
32 Expert systems Expert systems are used in many areas medical diagnosis, credit card authentication, process insurance claims, predicting weather, grammar checker in word processors, medical diagnosis, even prospectingThey have many advantagescan help to train new expertscan provide expertise when no experts are available physicallycan preserve knowledge of expertscan help to reduce the number of human errorscan augment human decision making when used collaboratively by an expert
33 Machine Learning Machine learning attempts to learn patterns from data various machine learning algorithms: eg. ID3data miningtext miningmining databasesusage miningneural networksMany applications now of machine learningin information retrievalin mining corporate databasesin modelling behaviour of users for personalized and adaptive systems
34 RobotsA robot is a computer-controlled machine designed to perform specific manual tasks.a robot’s central processor might be a microprocessor embedded in the robot’s shell, or it might be a supervisory computer that controls the robot from a distance
35 RobotsRobots are program-controlled machines designed to perform specific manual tasksremote controlled robots: intelligence is supplied by humans which control the robot from a distanceautonomous robots: the robot does its own reasoning, planning, and actingThey use input sensors to sense their current environment and then are able to act appropriately to this environment: planning, acting, and reactingThey can do things that are impossible for humanssee infrared lights, rotate arms 360°, enter areas unreachable by humans, etc.They are limited by current state of technologycan’t tie shoelace, can’t consistently tell the difference between a cat and dog!what is the algorithm for tying your shoelace??
36 Robots and planning Robots often need to be able to plan eg. a Mars robot decides how to go from one place to anothereg. a household robot decides how to avoid obstacles while vacuuming the floorA plan consists of a sequence of steps that the robot creates to carry out one or more tasksAdvantage of allowing the robot to plan: it can flexibly react to local conditionseg. Mars is a long way away, with many minutes to communicate back and forth; makes things very complex to control from eartheg. the dog keeps moving about during the vacuumingIt is useful to plan ahead of time but the planner must also be able to dynamically change a plan as it is executed when the unexpected happensPlanning is useful in many areas outside of roboticsinstructional planning in an e-learning systemdialogue planning in a conversational systemscheduling applications, eg. NASA has an intelligent scheduling system that has reduced the time to re-configure the shuttle by one third
37 Intelligent AgentsAgents are autonomous software processes that can reason and act in some environmentusually the environment is virtual (or else they would be called robots)agents in computer gamespersonal or companion agents in e-learning, eg. Stevesystem agentsMulti-agent systemsmore than one agent in the environmentbehaviour is emergent, based on agent interactionsthe environment and agents constitute an eco-system of a sortwhat if the agents could replicate? what if they could reason and communicate as well as humans? what would they talk about with each other? what would they tell us?
38 Are humans a multi-agent system? EnvironmentRules and constraintsSome interference from outside (God)The ant farm analogy
39 Implications of AIWeak AI: computers can be made to replicate some intelligent functions, but not allStrong AI: it will eventually be possible to create a computing system that equals human intelligence
40 Implications of AIWeak AI: computers can be made to replicate some intelligent functions, but not allwhich ones not?possibly emotions?Hollywood treats emotions as special: eg. insisting that emotions are uniquely human and all other entities wish to have them (such as Data on Star Trek: Next Generation)are emotions disconnected from cognition?others?the expertise paradox:we can make computers that are relatively good at doing things that require human expertisebut we have a very hard time capturing commonsense and basic abilities that children can perform at a young agewhy might this be?
41 Implications of AIStrong AI: it will eventually be possible to create a computing system that equals human intelligencelooking back: over history many things have been considered unique to humans but no longer: tool use; doing mathematics (originally “computer” meant a human doing calculations)why should any current human capabilities in principle be out of bounds for computers?Church’s thesis (paraphrased): there can be no more powerful system that processes information than a computera prospect with staggering implicationsif we can build a system that equals human capabilities, why not one that exceeds them?plenty of things that computers do better than humans now!probably more to come?are there any limits?towards “The Singularity”: human-machine symbiosis
42 Philosophical and Social Implications of AI How intelligent can computers get? Will they surpass humans?Should limits be imposed on what can be investigated?Who is responsible if the AI system makes a mistake?Are there tasks that computers should not even attempt? Tasks that only humans should do?How much would you trust a computer compared to a human?Will intelligent computers be infallible? - consider HAL in the movie 2001Can there be robot moral codes? - consider Asimov’s 3 laws of roboticsShould an AI of human-level intelligence (or above) have the same rights as humans? - consider David in the movie AICan intelligent programs have emotions? Should they? Can it be avoided?Would an artificial intelligence be immortal? - consider that it could save a copy of itself from time to time to be restored if anything happened to it. But, would a copy of an intelligent program be the same individual? What if the original copy was suddenly found again?Does the metaphor mind=software and brain=computer illuminate the mind-body problem in philosophy? Does it shed light on the notion of “soul”?
43 Asimov’s 3 laws of robotics Isaac Asimov, science fiction writer, introduced in his 1942 short story "Runaround", the Three Laws of Robotics:1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.2. A robot must obey orders given it by human beings except where such orders would conflict with the First Law.3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.How could these be enforced by the software running a robot?consider the Japanese engineer killed by an industrial robot pushing him into a grinding machine when it couldn’t sense him: how could this be prevented?
44 Resources to investigate expert system demoNouse perceptual vision interfacefledermausAiboAibo robot soccer playrobot car: desert challengerobot car: city challengerobomower
45 To know What is AI? What is the Turing Test? Two approaches to AI Natural language communicationPattern recognitionImage analysisOCRSpeech recognitionExpert systemsRobotsFuture of AI and ethical dilemmas