Unit 7: Early AI hits a brick wall

Transcription

1 Unit 7: Early AI hits a brick wall Language Processing ELIZA Machine Translation Setbacks of Early AI Success Setbacks Critiques Rebuttals Expert Systems New Focus of AI Outline of Expert Systems Assessment of Expert Systems copyright c 2013 Hyeong In Choi 1 / 29

2 Language Processing Language Processing ELIZA Machine Translation 2 / 29

3 Language Processing ELIZA Bobrow s STUDENT can be viewed as having some rudimentary natural language processing capability ELIZA: movie developed by J. Weizenbaum (1966) a program that simulated conversations with humans in ordinary (natural) language in reality, it has very limited vocabulary and has virtually no knowledge of actual human thought or language can be fooled very easily; brittleness 3 / 29

4 Language Processing Machine translation Attempt to translate written scientific/technical documents from Russian to English movie The resulting translation quality was deemed unacceptable 4 / 29

5 Setbacks of Early AI Setbacks of Early AI Success Setbacks Critiques 5 / 29

6 Setbacks of Early AI Success There have been many successes in AI movie in narrowly defined environment/context with limited, well-focused objectives Most successes are for problems/tasks that are by nature logical/symbolic like Logic Theorist SAINT Games(chess, checker, etc.) 6 / 29

7 Setbacks of Early AI Setbacks For problems/tasks requiring human-like intelligence with commonsense movie AI could not deliver the result whatever AI could achieve was contrived with special tricks For large realistic problems, the complexity gets too unwieldy too quickly heuristic methods frequently break down 7 / 29

8 Setbacks of Early AI Setbacks Lack of common sense movie The early AI products lack the knowledge of environment lack the sense of context are very brittle in short, have no common sense that all humans tacitly possess without ever being conscious about movie 8 / 29

9 Setbacks of Early AI Critiques Lots of criticism appeared. H.Dreyfus published a book What Computers Can t Do (1972) listed all the failures scathing attack not only on AI s achievements but on AI s very viability Lighthill report (1973) a systematic investigation on the contemporary state of art of AI BBC TV Debate (June,1973) ushered in the AI Winter 9 / 29

10 Rebuttals 10 / 29

11 Sir James Lighthill Lucasian Professor of Mathematics at Cambridge University investigated the contemporary state of art of artificial intelligence compiled a report commissioned by the British Science Research Council (SRC) in 1972 In June 1973, there was a BBC TV Debate centered on the Lighthill report with Lighthill, Michie, Gregory and McCarthy participating. 11 / 29

12 The report categorized the AI research into Category A: Advanced Automation Category B: Bridge Category/Building Robots Category C: Computer-based Central Nervous System (CNS) research 12 / 29

13 Category A: Advanced Automation Category A items written character recognition pattern recognition speech recognition and synthesis machine translation product design and assembly container packing exploration and action in hostile environments theorem proving inductive generalization analogy spotting information storage and retrieval analysis of chemical structures problem solving graph traversing learning and decision making 13 / 29

14 Objectives to replace human beings by machines for specific purposes look beyond automation done in control engineering and data processing aim to make a far fuller use of general-purpose digital computer s logical potentialities Longer term objectives combining a well structured knowledge-base and an advanced problem solving capability to generate improved methods for industrial and economic planning and decision making an increasingly greater capability of learning 14 / 29

15 Category C: Computer-based CNS research main focus theoretical investigations related to neurobiology and to psychology build models of CNS processes whose performance in due course can be compared with experimental data theories to interpret neurobiological data on specific areas of the CNS using computer-based models of neural nets to test out particular hypothesis construct computer models from which new theoretical concepts many develop Long-term aim understand human intellect 15 / 29

16 Category B: Bridge category/building Robots Bridge between Categories A and C Purports to be the coherent field of Artificial Intelligence research Building Robots seen as an essential bridge activity the concept of robot here is to mean automatic devices that mimic a certain range of human functions without seeking useful ends the research of this category is expected to feed into the work of Categories A and C 16 / 29

17 General aspects of work in Category B mimicking some special functions developed particularly in man co-ordination of eye and hand visual scene analysis use of natural language common-sense problem solving 17 / 29

18 Criticism: Assessment of achievement Categories A no breakthroughs some respectable achievements but disappointing when compared with the results of conventional control engineering and data processing subpar performance in pattern recognition character (printer/hand written) recognition speech recognition mathematical theorem proving notorious disappoints in machine translation severe criticism of 1966 US National Academy of Sciences study 18 / 29

19 Categories C no breakthroughs some benefits to psychology controversy over the role of neural network generally disappointing results category B biggest disappointment is in Category B doubts about whether the whole concept of AI as an integrated field of research is valid one the early enthusiasm for programming and building a robot that would mimic human ability in a combination of eye-hand co-ordination and common-sense problem solvings ended up gravely disappointed Lighthill expresses gravest doubt on the very viability of AI, especially in Category B 19 / 29

20 Why such failure? Combinatorial explosion data and relations (knowledge) explode exponentially algorithm must cope with exponentially exploding possibilities relies on heuristic methods to overcome this problem Problem with heuristic method heuristic methods tend to be ad hoc, relying on human knowledge/expertise heuristics vary greatly from situations to situations tantamount to adding specialized knowledge obtained through extremely careful study 20 / 29

21 common sense Lighthill did not explicitly single out the difficulty of representing (storing) common sense knowledge and the problem of handling it as one of the culprit but he was amply aware of the problem of common sense and mentioned about it in many places example: chess Lighthill mentioned chess playing machine was at best at the level of experienced amateur at that time despite effects of the past 25 years evaluation function is entirely due to human knowledge and skill intelligence is due to human computer offers only speed, reliability and biddability no learning chess program 21 / 29

22 Future possibilities (his personal view) Category A or C Some greater achievements possible this achievement will forge far stronger links to the immediate fields of application not much ties with Category B Category B will continue to be disappointing failure in Building Robots some robot s cognitive models on linguistics or problem-solving tasks will become integrated in Category C; while practical of engineering tasks will be integrated in Category A 22 / 29

23 Rebuttals Rebuttal by N.S.Sutherland Rebuttal by R.N.Needham Rebuttal by H.C.Longuet-Higgins Rebuttal by D.Michie Rebuttal by J.McCarthy 23 / 29

24 Expert Systems Expert Systems New Focus of AI Outline of Expert Systems Assessment of Expert Systems 24 / 29

25 Expert Systems New Focus of AI Ed Feigenbaum realized that the microworld paradigm cannot be extended to real, meaningful human activities but in a narrowly focused problem domain of specialists, it can yield results on par with human experts DENDRAL movie developed by Feigenbaum (1960s) chemical analysis program utilizing spectrographic information first successful expert system uses only a few hundred rules 25 / 29

26 Expert Systems New Focus of AI MYCIN claims that most of human specialist s knowledge on narrowly defined task can be reduced to a small number of rules developed by E.Shortliffe et al. (early 1970s) medical diagnostic system for identifying bacterially infected diseases recommending proper antibotics used around 600 rules used probability ranking 26 / 29

27 Expert Systems Outline of Expert Systems human expert dialog knowledge engineer human knowledge knowledge-base Facts Rules movie : Ontology and frames : relations,etc Reasoning Engine search inference logical resolution production system (rule engine) descriptive language (e.g., PROLOG) machine learning inexact/probabilistic reasoning 27 / 29

28 Expert Systems Assessment of Expert Systems Pros Cons flowering of experts systems in all kinds of specialized fields produced many useful results that are on par with human experts beginning of AI as an industry provoked lots of researches in mathematics, statistics and computer science brittle movie has no broader context in human life has no common sense 28 / 29

29 Expert Systems Assessment of Expert Systems mechanized reasoning difficult to cope with unforeseen events interior to human judgment in subtle situations acts like an idiot savant 29 / 29