Artificial Intelligence CPSC 327 Week 1 The Astonishing Hypothesis (with apologies to Francis Crick) 1 Towards a Definition • The name of the field is composed of two words: – Artificial • Art • Artifact • Artifice • Article – Intelligence What do these have in common? 2 So, AI is the construction of intelligent systems • “Artificial Intelligence (AI) may be defined as the • • • branch of computer science that is concerned with the automation of intelligent behavior.” p. 1 Key notion: behavior Classic def. of AI doesn’t care how it’s composed. Intelligent systems behave intelligently. 3 But what’s intelligence? 4 Some indicators • Ability to do mathematics • Ability to design a machine • Ability to play chess • Ability to speak • Ability to write an essay 5 What do all of these have in common? 6 AI has concentrated on those things that we get rewarded for in school. 7 A more precise set of criteria* 1. Intelligence must entail a set of skills to solve 2. genuine problems valued across cultures Potential isolation by brain damage: “to the extend that a particular faculty can be destroyed as a result of head trauma, its isolation from other faculties seems likely.” *Howard Gardner, Frames of Mind, Basic Books, 1993, pp. 62-67 8 Two more criteria 3. Existence of prodigies. That is, the skills may be plotted along a standard normal distribution. Some people are way out on the right side. 4. Existence of one or more basic information processing operations that deal with specific inputs. “One might go so far as to define a human intelligence as a neural mechanism or computational system which is genetically programmed to be activated or “triggered” by certain kinds of internally or externally presented information.” For example: – – – – – Sensitivity to pitch relations (musicians) Ability to see patterns among symbols (mathematicians) Ability to imitate bodily movements (athletes) Ability to understand emotional and power relations among a group of people (politicians) Ability to speak a language (all humans) 9 Two More 5. Evolutionary history and evolutionary plausibility. “A specific intelligence becomes more plausible to the extent that one can locate its evolutionary antecedents.” 6. Distinctive developmental history—levels of expertise through which every novice passes 10 Yet another 7. Support from experimental psychology. “To the extent that various specific computational mechanisms…work together smoothly, experimental psychology can also help demonstrate the ways in which modular … abilities may interact in the execution of complex tasks.” Psychometric findings are also relevant. 11 Finally 8. Susceptibility to encoding in a symbol system. “Much of human representation and communication … takes place via symbol systems—culturally contrived systems of meaning which capture important forms of information. Language, picturing, mathematics are but three of the symbol systems that have become important the world over for human survival and human productivity…Symbol systems may have evolved in just those cases where there exists a computational capacity ripe for harnessing….” 12 An Historical Aside • Newell & Simon, two AI pioneers, formulated the Physical Symbol System Hypothesis in their 1978 Turing Award Lecture: “A physical symbol system possesses the necessary and sufficient conditions for general intelligent action.” (about which, more later). 13 Gardner’s Seven Intelligences These 8 criteria lead to seven intelligences • Musical intelligence • Logical-mathematical intelligence • Linguistic intelligence • Spatial intelligence (kekule’ and the Benzene ring, artist) • Bodily-kinesthetic (athlete, dancer, surgeon) • Intrapersonal—access to one’s own emotional life (novelist, shaman) • Interpersonal—ability to read the emotional state of others (politician, gambler, therapist) . 14 The good and bad news • AI has had lots of success with logical intelligence • Less success with linguistic intelligence • Almost no success with what comes under the heading of common sense 15 Yet another definition • AI is the science of making machines do the sort of things that are done by human minds (Oxford Companion to Mind) • Why? I mean, who cares? 16 Five applications • Build various kinds of intelligent assistants – – – – Monitor email Perform hazardous tasks Monitor correct operations of a computer network Monitor/rewrite news • Make computers and other appliances easier to • • • use Machine translation Intelligent tutors Model human cognition 17 Model Human Cognition • Another Def. – AI is the study of mental faculties through the use of computational models 18 Good Points of this definition 1. Stays away from purely human intelligence by talking of mental faculties • • • • • Perceive the world Learn, remember, control action Create new ideas Communicate Create the experience of feelings, intentions, selfawareness 2. Introduces the notion of a computational model 19 Fundamental Assumption in AI • Computational/Representational Understanding of Mind – Theory can best be understood in terms of representational structures in the mind and computational procedures that act on them – Implication is that the material in which these are implemented is irrelevant 20 So • Material of the brain – Neural cells and electrical potential called synapses • Material of Computers – Silicon, copper, electrical impulses organized to implement the laws of symbolic logic 21 Central Feature of AI • Materials are irrelevant • Intelligence implemented in silicon is still intelligence • Turing Test laid out the ground rules over fifty years ago 22 Physical Symbol System Hypothesis • Allen Newell & Herbert Simon • “A physical symbol system has the necessary and sufficient means for general intelligent action.” • What is a PSS? – A program – A Turing Machine 23 To Explain • Symbol – May designate anything – If it designates something in the world, it has a semantics – May be manipulated according to rules and so has a syntax • Necessary – Any system that exhibits general intelligence, will prove, upon analysis, to be a physical symbol system 24 Further • Sufficient – Any physical symbol system of large enough size can be organized to exhibit general intelligent action • General Intelligent Action – Same scope as human behavior: in any real situation, behavior appropriate to the ends of the system and adaptive to the demands of the environment can occur 25 Example: Language Generation • Mary hit the ball. – Letters are symbols for sounds – Arranged according the rules of spelling – To form words – But, words refer to • Objects: Mary, John, Ball • Actions: hit • Relationships: to • These form the semantics of the sentence 26 • By arranging these words according to linguistic rules, called syntax, we get sentences • But how do we know the rules? • Language spoken by native speakers is data. Linguists tease out the regularities. • So, a grammar is descriptive, not prescriptive 27 Simple Context Free Grammar S NP VP VP V NP (PP) PP P NP NP (det) N det {a, the} N {Mary, John, ball, bat} P {to, with} V bat Try deriving the sentence: Mary hit the ball to John with the bat. Notice the recursive structure 28 So we have • Symbols • Syntax • Semantics If these were sufficiently complex, we would have a PSS that generates all English sentences. 29 The Astonishing Hypothesis • Intelligence is, at bottom, symbol manipulation • Convenient for computer scientists • Hard to know which came first – Claim then the computer – Computer then the claim • Western thought from Aristotle to Boole to Frege • has paid special attention to logic Especially interesting to learn that logic is pattern matching, a claim that I’ll argue for when we study proofs by resolution refutation 30 Objections/Counter Objections • Computers only do what they’re told – Debugging programs: we often don’t know what we’ve told computers to do – Rules given to AI program are like the axioms of an algebra. They allow the inference of the theorems that were not anticipated – PDP is not rule bound. Or at least, it’s difficult to specify the rules • Can’t specify rules to govern all of behavior – Machine learning • Searle’s Chinese box experiment • AI systems are brittle and not scaleable – PDP • Intelligence and logic are not the same thing – PDP – genetic algorithms – Hidden Markov models 31 AI Areas 1. Game playing – Source of results in state space search, state space representation, heuristic reasoning 2. Theorem Proving – Early successes: Theorem 2.85 from Principia – Problem: prove large number of irrelevant theorems before stumbling onto the goal 3. Expert systems – Domain-specific knowledge – Rigidly hand-crafted – Don’t learn Common threads to all three – Well-defined set of rules – No outside knowledge is required 32 4. NLP • Success with parsing • Success with speech synthesis and • • • transcription Growing success with translation All successes are probabilistic Language is deceptively rule-bound – – He saw her duck “janet needed some money. She got her piggy bank and shook it. Finally, some money came out.” • Why did Janet get the piggy bank? • Did Janet get the money? 33 • Why did Janet shake the piggy bank? 5. Cognitive Modeling • • Forces precision Existence proof 6. Robotics 7. Machine Learning (e.g., neural networks, evolutionary computing, stochastic models) 34 Two Strands in AI 1. – – – – Strand based on logic “The reliance on logic as a way of representing knowledge and on logical inference as the primary mechanism for intelligent reasoning are so dominant in Western philosophy that their “truth” often seems unassailable. It is no surprise, then, that approaches based on these assumptions have dominated the science of artificial intelligence from its inception to the present day.” p. 16 But various forms of philosophical relativism have questioned the “objective basis of language, science, and society” in the past half century. Examples come from philosophy of language (Wittgenstein, Grice, Austin, Searle), phenomenology (Husserl, Heidegger, Dreyfus), logic (Godel: In any logical system there must remain propositions that can’t be proven from within the system), linguistics (Winograd, Lakoff, usage-based linguists), post-modern thought (Derrida: “There is no outside the text”). The cumulative effect has been to call the AI project—at least as classically conceived—into question. 35 2. Strand based on biological metaphors and stochastic modeling – Artificial life and genetic algorithms take their inspiration from the principles of biological evolution. Intelligence as emergent. – Connectionism (PDP) takes it inspiration from a highly abstract view of neurons connected by synapses through a feedback mechanism – Hidden Markov models: a machine learning technique that makes Bayesian inferences for chains of events • Bayes Rule: P(X|Y) = (P(Y|X) * P(X))/P(Y)) • In English: the probability that we have class today given that today is Thursday equals the probability that today is Thursday given that we have class times the probability that we have class divided by the probability that today is Thursday. 36