MAS S66 New Des+na+ons in Ar+ficial Intelligence Goals and Direc+ons for Future Research
New Des3na3ons in Ar3ficial Intelligence
© Joscha Bach 2013 © Joscha Bach 2013
Welcome
• AI as a science vs. AI as engineering • What is intelligence? • Where should AI focus? • What is the right methodology for AI? • How can we measure progress? • AI and the cogni3ve sciences • Will academic AI disappear? • Present and future challenges • Underresearched areas, and terra incognita
9/14/15 FutureAI 2
© Joscha Bach 2013 © Joscha Bach 2013
The Million Dollar Ques5on
• If you had a spare Million $... What is the single most interes3ng AI ques3on you would put up as a challenge?
9/14/15 FutureAI 3
© Joscha Bach 2013 © Joscha Bach 2013
Organiza5on
• Seminar, 12 sessions • weekly mee3ngs, discussion & presenta3ons
Grading: presenta3on & short essay required • get in touch ASAP ([email protected]) • weight 1:2:2 (par3cipa3on, presenta3on, paper)
• structure & material is open to change and sugges3ons
9/14/15 FutureAI 4
© Joscha Bach 2013
Sessions: see futureai.media.mit.edu
9/21: Possibili3es for ar3ficial minds 9/28: The Lighthill debate: AI as engineering or AI as a science 10/5: Towards mapping contemporary AI. The Norvig/Chomsky debate 10/12: Columbus Day (no session) 10/19: Agents within agents. The Society of Mind
10/26: The Neocognitron and Deep Learning 11/2: Universal intelligence. From Solomonoff induc3on to AIXI 11/9: AI and Neuroscience 11/23: Affect and Mo3va3on
11/30: Measuring the Progress of AI. Benchmarks 12/7: Closing Discussion. Can we sketch a Map of Future AI research?
9/14/15 FutureAI 5
© Joscha Bach 2013 © Joscha Bach 2013
AI as a founda5onal cogni5ve science
• From psychophysics to cogni3ve science • Why computa3onal models? • Why not psychology? • Why not neuroscience? • Why not AI?
FutureAI 9/14/15 6
© Joscha Bach 2013 © Joscha Bach 2013
From Psychophysics to Cogni5ve Science
• Should the science of the mind be nomothe5c or descrip5ve?
à Psychophysics
9/14/15 FutureAI 7
Hermann von Helmholtz
Wilhelm Wundt Gustav Fechner
© Joscha Bach 2013 © Joscha Bach 2013
Psychoanalysis
9/14/15 FutureAI 8
© Joscha Bach 2013 © Joscha Bach 2013
Behaviorism
• McDougall, Pawlow • Watson, Skinner
9/14/15 FutureAI 9
© Joscha Bach 2013 © Joscha Bach 2013
Legacy of Behaviorism
• Defense against behaviorism • Liele focus on nonbehavioural aspects of cogni3on • Narrow experimental paradigms • Difficulty to formulate theories • Exclusive focus on human performance
9/14/15 FutureAI 10
© Joscha Bach 2013 © Joscha Bach 2013
Func5onalist theories of mind
Ar3ficial Intelligence:
• The mind is a computa3onal system • The mind is less than a Turing machine • Thinking, percep3on, feeling, voli3on, norma3vity, … have to be explained in terms of informa3on processing
• We can test our ideas on how the mind works by running them as computer programs
FutureAI 9/14/15 11
© Joscha Bach 2013 © Joscha Bach 2013
The Way to Ar5ficial Intelligence
• Logic, Automata: Peirce, Boole, Tarski, … • Computability: Turing, Church, Post, Gödel • Computers: Babbage, Zuse, von Neumann, … • Informa3on Theory: Shannon • Programming Languages: McCarthy, Rochester, … • Cyberne3cs: Wiener, Ashby, … • Symbol Systems: Newell, Simon, … • Neural Networks: McClelland, PiQs, … • Autonomous, situated Agents: Minsky, Brooks, …
FutureAI 9/14/15 12
© Joscha Bach 2013 © Joscha Bach 2013
Darthmouth Conference: 1956
Nathan Rochester
Marvin Minsky
John McCarthy
Claude Shannon
FutureAI 9/14/15 13
© Joscha Bach 2013 © Joscha Bach 2013
Turing 1950: Objec5ons against AI
1. The Theological Objec3on 2. The "Heads in the Sand" Objec3on
3. The Mathema3cal Objec3on 4. The Argument from Consciousness
5. Arguments from Various Disabili3es 6. Lady Lovelace's Objec3on 7. Argument from Con3nuity in the Nervous System
8. The Argument from Informality of Behaviour 9. The Argument from Extrasensory Percep3on
à Learning Machines • Build child-‐like machines that learn how to be intelligent
9/14/15 FutureAI 14
© Joscha Bach 2013 © Joscha Bach 2013
AI eras
• Turing 1950: Compu3ng Machinery and Intelligence
• The Low-‐Hanging Fruits: 1950ies/60ies • 1969: Connec3onism curbed • Expert Systems: 1970ies/1980ies • Connec3onism reloaded: 1980ies onwards • Late 1980ies: Nouvelle AI • Agents: 1990ies/00s • Currently: Sta3s3cal AI vs. Cogni3ve Systems
FutureAI 9/14/15 15
© Joscha Bach 2013 © Joscha Bach 2013
The AI Winter
• 1973: Lighthill Report
FutureAI 9/14/15 16
© Joscha Bach 2013 © Joscha Bach 2013
Lighthill Report
“Ar3ficial Intelligence: A general survey” A: Automa3on, robo3cs, applica3ons B: Building robots to understand cogni3on C: Close modeling of biology
9/14/15 FutureAI 17
© Joscha Bach 2013 © Joscha Bach 2013
Lighthill Report
“Ar3ficial Intelligence: A general survey” A: Automa3on, robo3cs, applica3ons B: Building robots to understand cogni3on C: Close modeling of biology
9/14/15 FutureAI 18
© Joscha Bach 2013 © Joscha Bach 2013
A new beginning
• Christopher Longuet-‐Higgins: “Cogni3ve Science”
• Alan Newell: “Unified Theories of Cogni3on”
• Newell and Simon: GPS • Cogni3ve Architectures: Soar, ACT-‐R, EPIC, … • Integra3on over all sciences of the mind
9/14/15 FutureAI 19
© Joscha Bach 2013 © Joscha Bach 2013
Main Premise of Cogni5ve Science
Mind as Machine
FutureAI 9/14/15 20
© Joscha Bach 2013 © Joscha Bach 2013
Mind as Machine
Percep3on, and what depends on it, is inexplicable in a mechanical way, that is, using figures and mo3ons. Suppose there would be a machine, so arranged as to bring forth thoughts, experiences and percep7ons; it would then certainly be possible to imagine it to be propor3onally enlarged, in such a way as to allow entering it, like into a mill. This presupposed, one will not find anything upon its examina3on besides individual parts, pushing each other— and never anything by which a percep3on could be explained. (GoYried Wilhelm Leibniz 1714)
FutureAI 9/14/15 21
© Joscha Bach 2013 © Joscha Bach 2013
Mind as Machine
Percep3on, and what depends on it, is inexplicable in a mechanical way, that is, using figures and mo3ons. Suppose there would be a machine, so arranged as to bring forth thoughts, experiences and percep7ons; it would then certainly be possible to imagine it to be propor3onally enlarged, in such a way as to allow entering it, like into a mill. This presupposed, one will not find anything upon its examina3on besides individual parts, pushing each other— and never anything by which a percep3on could be explained. (GoYried Wilhelm Leibniz 1714)
FutureAI 9/14/15 22
© Joscha Bach 2013 © Joscha Bach 2013
Main Premise of Cogni5ve Science
Mind as informa3on processing system: Computa7onalism
FutureAI 9/14/15 23
© Joscha Bach 2013 © Joscha Bach 2013
Computa5onalism
• Does encompass quantum compu3ng, too • Contemporary form of mechanism • Agnos3c wrt. materialist physicalism
• strong universal computa3onalism vs. strong cogni3ve computa3onalism vs. weak computa3onalism
9/14/15 FutureAI 24
© Joscha Bach 2013 © Joscha Bach 2013
Computa5on: Lambda Calculus
Alonzo Church, 1936
λ y.x(yz) ab Expression:= Variable | Function|(Expression)| Expression Expression Variable:= { a...z } Function:= λ Variable . Expression
FutureAI 9/14/15 25
© Joscha Bach 2013
Computa5on: Turing Machine
• Alan Turing (1936)
FutureAI 9/14/15 26
© Joscha Bach 2013 © Joscha Bach 2013
Problems with Turing Machines
• delivers wrong intui3ons • insufficient treatment of relevant details • minds can do less than Turing Machines • constraints on implementable Turing Machines don’t match biological constraints
9/14/15 FutureAI 27
© Joscha Bach 2013 © Joscha Bach 2013
Computa5onal Cogni5ve Models
• In reality: No problems with computa3on (only plays role in discussions between scep3cs and proponents)
• Problems of Neuroscience: descrip3ve, below level of
informa3on processing • Problems of Psychology: a-‐theore3c experimentalism • Problems of Philosophy: no tools for tes3ng theories • Problems of AI: methodology and goals
9/14/15 FutureAI 28
© Joscha Bach 2013 © Joscha Bach 2013
Challenges for AI
• Re-‐integrate Cogni3ve Sciences • Construc3onist methodology: implementa3on to make complex theories testable
• Focus on broad models
• Science instead of engineering
9/14/15 FutureAI 29
© Joscha Bach 2013
Sessions: see futureai.media.mit.edu
9/21: Possibili3es for ar3ficial minds 9/28: The Lighthill debate: AI as engineering or AI as a science 10/5: Towards mapping contemporary AI. The Norvig/Chomsky debate 10/12: Columbus Day (no session) 10/19: Agents within agents. The Society of Mind
10/26: The Neocognitron and Deep Learning 11/2: Universal intelligence. From Solomonoff induc3on to AIXI 11/9: AI and Neuroscience 11/23: Affect and Mo3va3on
11/30: Measuring the Progress of AI. Benchmarks 12/7: Closing Discussion. Can we sketch a Map of Future AI research?
9/14/15 FutureAI 30