Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Artificial general intelligence (AGI) “building thinking machines” © 2007 General Intelligence Research Group AGI vs “narrow” AI • examples of narrow AI: – face recognition – spam filtering – data mining – Google Common objections • • • • • • intelligence is not well-defined it’s too hard computing power is not there yet no unifying theory of AI we don’t understand the brain etc… All this is bull shit! AI pioneers • Alan Turing (1912-1954) • John von Neumann (1903-1957) John McCarthy (1927-) Marvin Minsky (1927-) Implications of AGI • complete automation • ethical issues • “Technological Singularity” Vernor Vinge (1944-) Ray Kurzweil (1948-) Representative AGI projects • • • • • • • • Cyc Soar, ACT-R Polyscheme LIDA SNePS AIXI OSCAR NARS • • • • • • Novamente Cog CAM-Brain HTM SAIL a2i2 and many more…. (listed by Pei Wang) Cyc • most-funded AI project in history ($10s of millions) • based on predicate logic • complete ontology • millions of facts, concepts Doug Lenat (1950-) Soar • Allen Newell (1927-1992) John E Laird • based on production rules & rete algorithm • learning – “chunking” Novamente • Ben Goertzel (1966-) • probabilistic logic based on “uncertain probabilities” • graph-based knowledge representation • genetic algorithms for learning • robot living in virtual reality • 2007 book: Artificial General Intelligence NARS • • • • Non-Axiomatic Reasoning System Pei Wang can learn from experience work with insufficient knowledge and resources • unified cognition: reasoning, learning, planning, etc… • 2006 book: Rigid Flexibility SNePS • • • • Semantic Network Processing System Stuart C Shapiro extends first-order logic belief revision / assumption-based truth maintenance • natural language understanding AIXI • Marcus Hutter • highly abstract • based on Kolmogorov complexity theory • KC is incomputable • learning may take forever! Polyscheme • Nick Cassimatis • integrates multiple methods of representation, reasoning, and problem-solving • procedural substrate • not “one model” CAM-brain • • • • • Hugo de Garis (1947-) neural network evolvable hardware cellular automata currently at Wuhan University SAIL • John Weng neural network-based • navigates and learns from environment autonomously Jeff Hawkins (1957-) • inventor of “Palm Pilot” • founded Redwood Neuroscience Institute • 2005 book: On Intelligence • HTM (Hierarchical Temporal Memory) • neurally-inspired Braininspired AI visual cortex Wiring of 6-layer cortex Neurally-inspired AI • feedforward neural network • Jeff Hawkins’ approach • problem: invariant recognition: translation, rotation, scaling Statistical learning • • • • takes place in a vector space requires many examples target = manifold difficult to learn concepts with variables eg: On(apple,table), On(car,road), etc… “Spatial” pattern recognition ANN, SVM, PCA, Clustering, etc… Logic-based vision • visual features logical representation Logical-vision example Quadrilateral() :∃e1:edge ∃e2:edge ∃e3:edge ∃e4:edge ∃v1:vertex ∃v2:vertex ∃v3:vertex ∃v4:vertex Connects(e1,v1,v2) ^ Connects(e2,v2,v3) ^ Connects(e3,v3,v4) ^ Connects(e4,v4,v1) “Syntactic” pattern recognition predicate logic formula: featurei relation1(feature1, feature2, …) ^ relation2(feature3, feature4, …) ^ … Spatial interpretation? Logic-based AI Avoid reinventing the wheel! Logic-based AI • first-order predicate logic (Prolog) • common objections: “brittle” “rigid” “binary” “not numerical” “just a theorem prover” • probabilistic / fuzzy logic • non-deductive mechanisms eg: abduction, induction Modules • • • • • • • perception (eg vision) pattern recognition inference natural language learning truth maintenance planning Architecture Pattern recognition • “neural characteristics” “soft computing” • Prolog: chair(X) :- leg1, leg2, leg3, leg4, seat, back, horizontal(seat), vertical(back),... leg1 leg2 leg3 leg4 … ... fuzzy values chair Pattern recognition – “chairs” more chairs still more chairs Pattern recognition • how humans recognize “concepts”? • [Michalski 1989] “2-tiered approach” rule-based vs instance-based • Prolog: chair :- chair1 chair :- chair2 chair :- chair3 ... chair :- (rule for general chair) Probabilistic logic • classical resolution [JA Robinson 1965] • Bayesian networks [eg Judea Pearl] Resolution algorithm • try to resolve formulas repeatedly until no more can be resolved PVQ ~P V R QVR Bayesian network • propositional First-order Bayes net • [Peter Norvig & Stuart • [David Poole 2003] Russell 2003] • [Manfred Jaeger 1997] • [Kathryn B Laskey 2006] etc… BeltStatus(belt) RoomTemp(room) EngineStatus(machine) Bayesian vs classical logic • Conditional Probability Table (CPT) classical Bayesian (A ^ B) A B C A T T F F B T F T F C 1.0 0.0 0.0 0.0 A T T F F B T F T F C T F F F KBMC • Knowledge-Based Model Construction • [Wellman et al 1992] • generate Bayesian networks “on-the-fly” to answer specific queries KB KBMC example KBMC example Belief bases vs belief sets • belief set = Cn( belief base ) set of consequences • belief sets are too large to manipulate • for AGI, must use belief base Fuzzy logic • “John’s girl friend is probably very pretty” • fuzziness probability • Lotfi Zadeh (1921-) 1965 fuzzy sets 1973 fuzzy logic Confidence • Example: A. 10 girls, 5 have long hair B. 1000 girls, 500 have long hair p = 0.5 but A and B are not the same B has higher confidence • used in Pei Wang’s NARS logic Probabilistic-fuzzy inference ( P, C, Z )n ( P, C, Z ) x1 x2 ... probability confidence fuzziness Ps and Zs can be point-valued or interval-valued Probability intervals • Example: marry fool [p = 0.8] ! marry loser [p = 0.7] p( fool V loser ) = 0.7 + 0.1 * p( marry ) [ 0.7, 0.8 ] unknown Conditional probability table (CPT) • All permutations of fuzzy values • Or, store in a “distribution-free” format? a z1 z2 z3 z4 … b … … … … … C (P1, C1, Z1) (P2, C2, Z2) (P3, C3, Z3) (P4, C4, Z4) … “Rules of thought” • “If cats have claws, and Juney is a cat, then Juney has claws.” • P,x,y P(x) ^ isa(y,x) P(y) • modus ponens: P, P Q Q • syllogisms reasoning deduction induction retroduction abduction Abduction • “finding explanations” • eg glass is wet it was raining • algorithm: reverse of deduction (eg resolution) • very high complexity 0 (within the arithmetical complexity class 2 ) Abduction algorithm Induction vs abduction • abduction: answer = ground literals eg “grass is wet” “it was raining” • induction: answer = general formulae eg daughter(X,Y) :- father(Y,X) ^ female(Y) Induction • learning general patterns statistically • ILP (Inductive Logic Programming) [Stephen Muggleton] 1990s Induction example Given data: male(mary) = false female(mary) = true mother(mary, louise) = true father(mary, bob) = true daughter(bob, mary) = true daughter(X,Y) :- father(Y,X) ^ female(Y) Natural language • unifying framework • language = knowledge-based inference • [Jerry R Hobbs] “Abduction as Interpretation” eg “The Boston office called.” • “apple pie” “door knob” “street hawker” • all we need is a lot of rules • can inductively learn the rules Belief maintenance • • • • Truth Maintenance System (TMS) belief revision to attain “consistency” avoid “cognitive dissonance” Truth maintenance justifications Belief revision • “Epistemic entrenchment” • [Mary-Anne Williams 1995] Belief Base … … 6 5 4 entrenchment ranking 3 2 1 0 “Click” feeling Perhaps an effect of successful inference, abduction, or belief revision? Paraconsistency • holding 2 contradictory beliefs in the knowledge base at the same time Associative memory • • • • knowledge base = database special indexing to allow associative recall hard disk = long-term memory RAM = working memory Planning Conclusions • “neural” is problematic • “blank slate” is problematic • “logic-based” is very promising Agenda for Logic-based AI 1. design probabilistic-fuzzy logic 2. develop algorithms for: – abduction – belief maintenance 3. acquire common sense knowledge “Web 2.0”-style collaboration • • • • branching voting commercial problem: too few members Thank you • [Aliseda 2006] Abductive Reasoning: Logical Investigations into Discovery and Explanation. Synthese Library Series vol 330, Springer • [Antoniou 1997] Nonmonotonic Reasoning, MIT Press • [Cussens 2001] Integrating probabilistic and logical reasoning. In David Corfield and Jon Williamson eds Foundations of Bayesianism, volume 24 of Applied Logic Series, pages 241-260. Kluwer, Dordrecht [2000 Flach & Kakas eds] Induction and Abduction, Springer Applied Logic Series #18 [Haddawy 1994] Generating Bayesian networks from probability logic knowledge, in Proceedings of the 10th conference on uncertainty in AI, 1994. [Hobbs 200?] Abduction as Interpretation [Jaeger 1997] Relational Bayesian networks. In Proceedings of the 13th Annual Conference on Uncertainty in AI (UAI-97), p266-273, San Francisco, CA, 1997, Morgan Kaufman Publishers [Kakas, Kowalski, Toni 1992] Abductive Logic Programming, Journal of Logic and Computation 2(6):719-770. http://citeseer.ist.psu.edu/kakas93abductive.html [Laskey 2006] MEBN: A logic for open-world probabilistic reasoning. GMU C4I Center Technical Report C4I-06-01. George Mason Univ, USA. [Milch & Russell 2007] First-Order Probabilistic Languages: Into the Unknown In ILP: • • • • • • • Proceedings of the 16th International Conference on Inductive Logic Programming. Berlin: Springer • [Michalski 1989] Two-tiered concept meaning, inferential matching, and conceptual cohesiveness. In Vosniadou & Ortony eds, Similarity and analogical reasoning, p122-145. Cambridge University Press, New York. • [Muggleton 1996] Stochastic logic programs. In de Raedt, ed, Advances in Inductive Logic Programming, p254-264, IOS Press 1996. • [Ngo, Haddawy, & Helwig 1995] A theoretical framework for contextsensitive temporal probability model construction with application to plan projection. In Proceedings of the 11th Annual Conference on Uncertainty in Artificial Intelligence (UAI-95), p419-426, Montreal, Quebec, Canada. • [Norvig & Russell 2003] Artificial Intelligence: A Modern Approach, Prentice Hall. • [Poole 1993] Probabilistic horn abduction and Bayesian networks, Artificial Intelligence, 64(1), 81-129, 1993 • [Poole 2003] First-order probabilistic inference, Proc, IJCAI-03, Acapulco, August 2003, p985-991 • [Wellman, Breese, Goldman 1992] From knowledge bases to decision models. Knowledge Engineering Review 7(1): 35-52 • [Williams 1995] Changing nonmonotonic reasoning inference relations, in Proceedings of the second world conference on the fundamentals of AI, 469-482, Ankgor, Paris, 1995