AI History CE417: Introduction to Artificial Intelligence Sharif University of Technology Spring 2012
Ancient History The intellectual roots of AI and intelligent machines (human-like artifacts) in mythology Real (and fraudulent) mechanical devices behaving with some degree of intelligence.
Modern History By emerging modern computers, it became possible to create programs performing difficult intellectual tasks. From these programs, general tools are constructed which have applications in a wide variety of everyday problems. Emerging computing programmable devices (electronic computers) was a major breakthrough to make intelligent systems.
AI Timeline Early Successes Predictions that AI would eventually do almost anything 1943 McCulloch & Pitts: Boolean circuit model of brain 1950 Turing's "Computing Machinery and Intelligence paper 1956 Dartmouth meeting: "Artificial Intelligence" term coined 1952-69 Early AI progress, great expectations 1965 Robinson's complete algorithm for logical reasoning 1966-73 AI discovers computational complexity Neural network research almost disappears 1969-79 Early development of knowledge-based systems 1980-- AI becomes an industry 1986-- Neural networks return to popularity 1987-- AI becomes a scientific method 1995-- The emergence of intelligent agents 2001-- AI on very large datasets Dark Age Crawl back Industrial & Scientific Age
Periods in AI (briefly) Early period - 1950 s & 60 s (mostly based on search) Game playing (brute force), theorem proving (symbol manipulation), biological models (neural networks) Symbolic application period - 70 s Early expert systems, use of knowledge Commercial period - 80 s knowledge/ rule bases Scientific & Industrial period - 90 s and early 21 st Century Rapid advance due to greater use of solid mathematical methods and rigorous scientific standards Real-world applications
The Gestation of AI (1943-1956) Neural Network The first AI work: Modeling of Neurons Warren McCulloch & Walter Pitts, 1943 Any computable function could be computed by some network of connected neurons Learning neural network (Hebbian rule): updating rule for modifying the weights of connection between neurons Donald Hebb, 1949 First neural network computer (SNARC) Marvin Minsky & Dean Edmonds (undergraduate students at Harvard), 1950 Minsky studied universal computation in neural networks during his PhD at Princeton Later, Minsky proved theorems showing limitations of NN
The Gestation of AI (1943-1956) Turing Alan Turing (1950) Computing Machinery and Intelligence (1950) paper includes a complete vision of AI Turing introduced the Turing test, machine learning, genetic algorithms, and reinforcement learning fields First Chess Player Program Claude Shannon & Alan Turing, 1950s
The birth of AI (1956) John McCarthy organized a 2 month workshop at Dartmouth College McCarthy (Stanford), Minsky (MIT), Simon & Newell (CMU), Samuel (IBM) every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. Achieved no new breakthroughs but AI was dominated by these people and their students and colleagues for the next 20 years Artificial Intelligence name was chosen by McCarthy during workshop Why AI becomes a separate field: AI duplicates human faculties like creativity, self-improvement, and language use Methodology: a branch of computer science and the only filed trying to build machines functioning autonomously in complex, changing environments Newell and Simon from CMU presented the most general program Logic Theorist (LT) as a reasoning program (proved many mathematical theorems)
Early enthusiasm, great expectations (1952-1969) - Look, Ma, no hands! Many successes (in a limited way) in early years of AI In few years computers from doing just arithmetic to machines did anything remotely clever General Problem Solver (GPS) CMU (Simon & Newell, 1960) Imitated human thinking Geometry Theorem Prover IBM (Gelenter, 1959) proved theorems that many students of mathematics would find tricky Checkers Player Machines (Arthur Samuel, 1952) Using game tree search & Reinforcement Learning McCarthy, MIT, 1958 LISP, Time Sharing, Advice Taker (the first complete AI system)
Early enthusiasm, great expectations (1952-1969) - Look, Ma, no hands! McCarthy (logic) vs. Minsky (anti-logical outlook) Minsky s group chose limited problems known as microworlds appeared to require intelligence to solve. e.g. closed form calculus integration problems, geometric analogy problems that appear in IQ tests, blocks world NN of McCulloch-Pitts flourished Enhancing learning by Widrow (1960, 1962) rules Perceptron by Rosenblatt (1962) and convergence theorm
A dose of realty (1966-1973) Herbert Simon, 1957 The power of A.I. will increase so rapidly that in a visible future, the range of problems they can handle will be coextensive to that of human. Predictions came true within 40 years instead of 10 years Problems (Early systems turned out to fail on wider selections or more difficult problems) Most of early programs contained little or no knowledge of subject matter (did only simple syntactic manipulations) 1966, There is no Machine Translation for general scientific text and there would be no in immediate prospect. Intractability of problems ( Combinatorial Explosion ) Failed to prove theorems involving more than a dozen of facts Lighthill report, 1973Cancellation of all AI research in G.B. except to two universities Fundamental limitations on basic structures used to generate intelligent behavior Perceptrons although they learn anything that they could represent, they could represent very little
Knowledge based systems: The key to power (1969-1979) First decade of AI research General purpose search mechanisms (weak methods general but cannot scale up) Alternative more powerful, domain specific knowledge DENDRAL, 1969 - Inferring molecular structure MYCIN, 1971 - Diagnosis of blood infections with 450 rules Natural language understanding Shrdlu Blocks world Schank, Yale Demands for workable knowledge representation schemes (Prolog, PLANNER, Minsky s idea of frames)
AI Becomes industry 1980-present R1 Expert System at DEC, 1982 Configure orders for new computer systems Saving $40 million per year The Fifth Generation Project, 1981 (Japanese) 10 year plan to build intelligent computers running Prolog Counter attacks in U.S. and G.B. From a few million dollars in 1980 to billions of dollars in 1988 Expert systems, vision systems, robots, software and hardware specialized for these purposes
The return of neural networks 1986-present Reinvention of BACK-PROPAGATION First in 1969, then in 1986. Connectionist Connectionist vs. Symbolic Symbolism: manipulating knowledge of the world as explicit symbols (e.g., words), where these symbols have clear relationships to entities in the world Connectionism: embodying knowledge by assigning numerical conductivities or weights to connections inside a network of nodes
AI adopts the scientific method 1987- present It is more common to build on existing theories than to propose brand-new ones To base claims on rigorous theorems (rather than intuition) and hard experimental evidence (real applications rather than toy examples) Early isolation of AI from the rest of computer science has been abandoned (Neats defeated Scruffies) Samples of revolutions HMM for speech recognition and machine translation Baysian network for uncertain knowledge representation and reasoning NN became comparable to corresponding techniques (e.g. statistics)
Emergence of Intelligent Agents 1995-present Whole Agent Reorganizing previously isolated subfields of AI Influential founders of AI have expressed discontent with the progress of AI AI should put less emphasis on creating ever-improved version of applications that are good at a specific task AI should return to its roots machines that think, that learn, and that create (Human-level AI or HLAI) Artificial General Intelligence (AGI), 2007 Universal algorithm for learning and acting in any environment
Large Data Sets 2001-present Data became more important than algorithm Word-sense disambiguation Performance increasing yield from using more data exceeds any difference in algorithm choice Filling in holes of a photograph Poor when 10000 photos available while excellent when 2000000 photos in collection Knowledge bottleneck Learning with enough data instead of hand-coded knowledge engineering
Some samples of AI Successes Game playing: Deep Blue defeated Kasparov (1997) Problem solving A mathematical conjecture (Robbins conjecture) unsolved for decades was proved (1996) Planning NASA's autonomous planning program controlled the scheduling of operations for a spacecraft US forces deployed an AI logistics planning and scheduling program DART that involved up to 50,000 vehicles, cargo, and people transportation (During 1991 Gulf War) Robotics & robot vehicles NASA AI agent ran a satellite beyond Mars for over a day, without ground control (1999) Sojourner, Spirit, and Opportunity explore Mars NASA Remote Agent in Deep Space I probe explores solar system DARPA grand challenge: Autonomous vehicle navigates across desert and urban irobot Roomba automated vacuum cleaner, and PackBot used in Afghanistan and Iraq wars Speech understanding systems for airline Spam filters using machine learning Question answering systems automatically answer factoid questions Usable machine translation by Google
Which of the following can be done at present? Play a decent game of table tennis Drive in the center of Cairo Buying a week s worth of groceries at the market. Play a decent game of bridge Discover and prove a new mathematical theorem Write an intentionally funny story Give competent legal advice in a specialized area of law Translate spoken English into spoken Swedish in real time