Transcript INTRODUCTION TO ARTIFICIAL INTELLIGENCE
INTRODUCTION TO ARTIFICIAL INTELLIGENCE
Massimo Poesio LECTURE 1: Intro to the course, History of AI
ARTIFICIAL INTELLIGENCE: A DEFINITION • • • The branch of Computer Science whose aim is to develop machines able to display intelligent behavior Strong AI: developing machines all but undistinguishable from human beings Weak AI: developing SIMULATIONS of human intelligence
AI BY EXAMPLE: GAMES
AI BY EXAMPLE: ROBOCUP
AI BY EXAMPLE: LANGUAGE
AI BY EXAMPLE: LANGUAGE
A BRIEF HISTORY OF AI
• • • • • • Forerunners, I: logic and ontologies Forerunners, II: mechanical machines / robots The beginning of AI: Turing, Dartmouth, Games, Search The role of Knowledge in Human Intelligence The role of Learning Modern AI
ARISTOTLE
• Aristotle developed the first theory of knowledge and reasoning – his ideas eventually developed into modern – LOGIC – ONTOLOGIES
ARISTOTLE: SYLLOGISM (Prior Analytics) The first attempt to develop a precise method for reasoning about knowledge: identify VALID REASONING PATTERNS, or SYLLOGISMS BARBARA: A: All animals are mortal A: All men are animals.
A: Therefore, all men are mortal.
DARII: A: All students in Fil., Logica & Informatica take Intro to AI I: Some students in Filosofia take Fil., Logica & Informatica.
I: Therefore, some student in Filosofia takes Intro to AI.
FIRST ONTOLOGIES
LOGIC: BEYOND ARISTOTLE
• • • • Ramon Lull (13 th Century): first mechanical devices for automatic reasoning (Lull’s disks) Leibniz (17 th Century): Encoding for syllogisms Boole (19 th Century): Boolean Algebra Frege (1879): Predicate calculus
BOOLEAN ALGEBRA
FORERUNNERS, 2
TURING
• • • Alan Turing is the father of AI He was the first to imagine machines able of intelligent behavior And devised an intelligence test, the TURING TEST: replace the question “Can a machine be endowed with intelligence” with the question: – Can a machine display such human-like behavior to convince a human observer that it is a human being?
DARTMOUTH
• In 1956 a group of researchers including J. McCarthy, M. Minsky, C. Shannon, N. Rochester organized a workshop at Dartmouth to study the possibility of developing machine intelligence
THE BEGINNINGS OF AI (1956-1966) • • Early AI researchers identified intelligence with the kind of behavior that would be considered intelligent when displayed by a human, and tried to develop programs that reproduced that behavior Examples: – Chess – Theorem proving
HEURISTIC SEARCH
• This early research focused on the development of SEARCH ALGORITHMS (A*) that would allow computers to explore a huge number of alternatives very efficiently
THE SUCCESS OF EARLY AI
In 1997, the chess-playing program DEEP BLUE developed by IBM researchers led by Feng-hsiung Hsu , beat the chess world champion Gary Kasparov over six games
EARLY AI RUNS INTO TROUBLE (1966 1973) • Soon however researchers realized that these methods could not be applied to all problems requiring intelligence, and that there were a number of ‘simple’ problems that could not be handled with these methods at all – Example of the first: machine translation (the ALPAC report) – Example of the second: natural language, vision
COMMONSENSE KNOWLEDGE IN LANGUAGE UNDERSTANDING • Winograd (1974): – The city council refused the women a permit because they feared violence.
– The city council refused the women a permit because they advocated violence
AI KEY DISCOVERIES, 1
• Performing even apparently simple tasks like understanding natural language requires lots of knowledge and reasoning
THE ‘KNOWLEDGE YEARS’ (1969-1985) • • • Development of knowledge representation techniques Development of EXPERT SYSTEMS Development of knowledge-based techniques for – Natural Language Understanding – Vision
KNOWLEDGE REPRESENTATION METHODS • • • • Logic is the older formalization of reasoning It was natural to think of logic as providing the tools to develop theories of knowledge and its use in natural language comprehension and other tasks Great success in developing theorem provers But AI researchers quickly realized that the form of logic required was not valid deduction
FROM LOGIC TO AUTOMATED REASONING • • Starting from the ‘50s AI researchers developed techniques for automatic theorem proving These techniques are still being developed and have been used to prove non-trivial theorems
RESOLUTION THEOREM PROVING
All Cretans are islanders. All islanders are liars. Therefore all Cretans are liars. ∀ X C(X) implies I(X) ∀ X I(X) implies L(X) Therefore, ∀ X C(X) implies L(X) ¬C(X) ∨ ¬I(Y) ∨ I(X) L(Y) ¬C(X) ∨ L(Y)
HIGH PERFORMANCE THEOREM PROVING • There are now a number of very efficient theorem provers that can be used to demonstrate sophisticated mathematical theorems – Otter – Donner & Blitzen
THE FOUR-COLOR PROBLEM
• • Conjecture: given a plane divided in regions, it is possible to color the regiones in such a way that two adjacent regions are always of different colors using no more than 4 colors This conjecture was demonstrated by an automatic theorem prover in 1997
AI KEY DISCOVERIES, 2
• Neither commonsense nor ‘expert’ reasoning involve only valid inferences from certain premisses: – Commonsense reasoning often involves jumping to plausible conclusions – Expert reasoning involves making decisions with uncertainty
COMMONSENSE KNOWLEDGE IN LANGUAGE UNDERSTANDING • Winograd (1974): – The city council refused the women a permit because they feared violence.
– The city council refused the women a permit because they advocated violence
DIAGNOSTIC REASONING IN EXPERT SYSTEMS IF patient has loss of weight AND anorexia AND spleen is palpable AND fever THEN patient MAY have Hodgkin’s disease
BEYOND VALID INFERENCE
• As a result of discovery number 2, AI researchers quickly got involved with developing better theories of commonsense reasoning and uncertain reasoning – Drawing inspiration from psychology (work on semantic networks, scripts) – Developing new inference formalisms
SOME RESULTS FROM COGNITIVE PSYCHOLOGY: SEMANTIC NETWORKS • • Collins & Quillian, 1969: knowledge appears to be organized around objects and in a taxonomic way – A canary is yellow – – A canary has feathers A canary eats food Haviland & Clark 1974, Sanford & Garrod 1979: ‘associated’ knowledge available when concepts are mentioned – – I looked around the house.
The lounge was very spacious.
AN EXAMPLE OF SEMANTIC NETWORK
COMMONSENSE REASONING: SCRIPTS (SCHANK & ABELSON, 1977) John entered the restaurant. The menu had changed, so he ordered one of the new dishes. The bill was very steep.
AN EXAMPLE OF SCRIPT
‘KNOWLEDGE-BASED’ SYSTEMS FOR LANGUAGE UNDERSTANDING: SHRDLU Winograd, 1974
A DIALOGUE WITH SHRDLU Person: PICK UP A BIG RED BLOCK.
Computer: OK.
Person: GRASP THE PYRAMID.
Computer: I DON'T UNDERSTAND WHICH PYRAMID YOU MEAN.
Person (changing their mind): FIND A BLOCK WHICH IS TALLER THAN THE ONE YOU ARE HOLDING AND PUT IT INTO THE BOX.
Computer: BY "IT", I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER THAN THE ONE I AM HOLDING.
Computer: OK.
Person: WHAT DOES THE BOX CONTAIN?
Computer: THE BLUE PYRAMID AND THE BLUE BLOCK.
Person: WHAT IS THE PYRAMID SUPPORTED BY?
Computer: THE BOX.
Person: HOW MANY BLOCKS ARE NOT IN THE BOX?
Computer: FOUR OF THEM.
Person: IS AT LEAST ONE OF THEM NARROWER THAN THE ONE WHICH I TOLD YOU TO PICK UP?
Computer: YES, THE RED CUBE.
EXPERT SYSTEMS
• Beginning with the ’70s, AI researchers developed systems that could be used to diagnose – Illnesses (MYCIN, DXPLAIN) – Mechanical problems • DELTA-CATS1, General Electric, Bonissone et al 1984 • IDEA, Centro Ricerche Fiat, ~1993 – etc
EXPERT SYSTEMS
A COMMONSENSE ENCYCLOPEDIA: CYC • • • A project initiated in 1984 by Doug Lenat. The goal: to encode all of commonsense knowledge Changed the underlying formalism several times. – These days: a logic-based representation Two versions available: – OpenCyc ( http://www.opencyc.org/ ) • 50 000 concepts, 300 000 facts • Can be downloaded / on the Web – ResearchCyc (http://research.cyc.com/) • 300 000 concepts, 3 million facts
KNOWLEDGE IN CYC
"Bill Clinton belongs to the class of US Presidents“ (#$isa #$BillClinton #$UnitedStatesPresident) “All trees are plants” (#$genls #$Tree-ThePlant #$Plant) “Paris is the capital of France". (#$capitalCity #$France #$Paris)
COMMONSENSE REASONING
• Modelling commonsense inference required the development of entirely new paradigms for inference beyond classical logic • Non-monotonic reasoning • Probabilistic models
AI RUNS INTO TROUBLE, AGAIN
• • The CYC project started in 1984, and by common opinion is nowhere near finished – – Hand-coding of commonsense FACTS is unfeasible (We will get back to this point later when talking about socially constructed knowledge) Work on lower-level tasks such as speech perception revealed the impossibility of hand coding commonsense RULES and assigning them priorities
SPEECH
KEY AI DISCOVERIES, 3
• A theory of intelligence requires a theory of how commonsense knowledge and cognitive abilities are LEARNED
THE MACHINE LEARNING YEARS (1985-PRESENT) • • The development of methods for learning from evidence started even before Dartmouth But machine learning has now taken center stage in AI
CYBERNETICS
• McCulloch, Pitts (1943): first artificial neurons model (based on studies of real neurons)
KNOWLEDGE REPRESENTATION IN THE BRAIN
MODELS OF LEARNING BASED ON THE BRAIN: THE PERCEPTRON
LEARNING TO CLASSIFY OBJECTS
ARTIFICIAL INTELLIGENCE TODAY
• • Artificial intelligence as a science: – Artificial intelligence vs. Cognitive Science Artificial Intelligence as a technology
AI INDUSTRY: GOOGLE
AI INDUSTRY: ROBOTICS
CONTENTS OF THE COURSE
• • • Knowledge Representation – A reminder about logic – – Ontologies Semantic networks Machine Learning – A reminder about statistics – Supervised learning – Unsupervised learning Putting it all together: Natural Language – A task that requires to use both knowledge representation and machine learning
PRACTICAL INFORMATION
• • • • • 60 hours / 12 credits Timetable: – Mondays, 10-12 and 16-18 (lectures) – Tuesdays, 12-14 (labs / tutorials) Prerequisites – The ideal student would have the background provided by the three-year course in Filosofia and Informatica (some background in linguistic, logic, and statistics; some experience with programming) Evaluation – A project to be presented at the exam Web site: http://clic.cimec.unitn.it/massimo/Teach/AI
READING MATERIAL
• • • Required: – The course slides, available from the Web Site – Other material downloadable from the Website Recommended readings: – Russel and Norvig, Artificial Intelligence: A Modern Approach (2 nd ed), Prentice-Hall – Bianchini, Gliozzo, Matteuzzi, Instrumentum vocale: intelligenza artificiale e linguaggio, Bononia Supplementary on specific sub-areas of AI: – KR: • • John F. Sowa, Knowledge Representation, Brooks / Cole Blackburn, Bos, Representation and Inference for Natural language, CSLI – ML: • Mitchell – Machine Learning – Prentice-Hall
READINGS
• This lecture: – http://en.wikipedia.org/wiki/Timeline_of_artificial_intelligence – John F. Sowa, Knowledge Representation, Brooks / Cole, ch. 1