The Artificial Jack of All Trades: The Importance of Generality in - - PowerPoint PPT Presentation

the artificial jack of all trades the importance of
SMART_READER_LITE
LIVE PREVIEW

The Artificial Jack of All Trades: The Importance of Generality in - - PowerPoint PPT Presentation

The Artificial Jack of All Trades: The Importance of Generality in Approaches to AI Tarek R. Besold KRDB, Faculty of Computer Science, Free University of Bozen-Bolzano 29. October 2015 Tarek R. Besold The Importance of Generality in AI Honor


slide-1
SLIDE 1

The Artificial Jack of All Trades: The Importance of Generality in Approaches to AI

Tarek R. Besold

KRDB, Faculty of Computer Science, Free University of Bozen-Bolzano

  • 29. October 2015

Tarek R. Besold The Importance of Generality in AI

slide-2
SLIDE 2

Honor to whom honor is due...

The following is joint work with: Ute Schmid Faculty of Information Systems and Applied CS, University of Bamberg

Tarek R. Besold The Importance of Generality in AI

slide-3
SLIDE 3

Intelligence, Cognition, and Computer Systems

Tarek R. Besold The Importance of Generality in AI

slide-4
SLIDE 4

Intelligence, Cognition, and Computer Systems

Nilsson (2009) AI is that science devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment. McCarthy (Dartmouth proposal, 1956) The study is to proceed on the basis of the conjecture that 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.

Tarek R. Besold The Importance of Generality in AI

slide-5
SLIDE 5

Intelligence, Cognition, and Computer Systems

Nilsson (2009) AI is that science devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment. McCarthy (Dartmouth proposal, 1956) The study is to proceed on the basis of the conjecture that 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.

Tarek R. Besold The Importance of Generality in AI

slide-6
SLIDE 6

Intelligence, Cognition, and Computer Systems

Standard (“weak”) AI Computational models can be used to simulate human information processes thereby either providing tools which take over specific functions or tasks previously requiring certain mental capacities, or allowing detailed and consistent generative descriptions of specific sub-areas of cognition. Human-level AI The (re)creation of human higher-level cognitive or intellectual capacities by artificial means is possible and will eventually be achieved by scientific means.

Tarek R. Besold The Importance of Generality in AI

slide-7
SLIDE 7

Intelligence, Cognition, and Computer Systems

Standard (“weak”) AI Computational models can be used to simulate human information processes thereby either providing tools which take over specific functions or tasks previously requiring certain mental capacities, or allowing detailed and consistent generative descriptions of specific sub-areas of cognition. Human-level AI The (re)creation of human higher-level cognitive or intellectual capacities by artificial means is possible and will eventually be achieved by scientific means.

Tarek R. Besold The Importance of Generality in AI

slide-8
SLIDE 8

Intelligence, Cognition, and Computer Systems

Working hypothesis While standard AI can confine itself to the modelling and (mostly descriptive) study of individual mental capacities as isolated subparts

  • f the mind, HLAI necessarily has to take a general holistic

interpretation of intelligence and cognition as foundation. Challenges for HLAI approach What is “intelligence” in the first place? How can intelligence be detected and how can HLAI systems be evaluated? How can progress be measured?

Tarek R. Besold The Importance of Generality in AI

slide-9
SLIDE 9

Intelligence, Cognition, and Computer Systems

Working hypothesis While standard AI can confine itself to the modelling and (mostly descriptive) study of individual mental capacities as isolated subparts

  • f the mind, HLAI necessarily has to take a general holistic

interpretation of intelligence and cognition as foundation. Challenges for HLAI approach What is “intelligence” in the first place? How can intelligence be detected and how can HLAI systems be evaluated? How can progress be measured?

Tarek R. Besold The Importance of Generality in AI

slide-10
SLIDE 10

Computers, Minds, Intelligence

Tarek R. Besold The Importance of Generality in AI

slide-11
SLIDE 11

Computers and the Mind

Usual theoretical foundations at the heart of many endeavors in HLAI and/or computational cognitive modeling:

1

“Computer metaphor” of the mind (i.e. the concept of a computational theory of mind).

2

Church-Turing thesis.

1

Bridges gap between humans and computers:

Human mind and brain can be seen as information processing system. Reasoning and thinking corresponds to computation as formal symbol manipulation.

2

Gives account of the nature and limitations of the computational power of such a system.

Tarek R. Besold The Importance of Generality in AI

slide-12
SLIDE 12

Computers and the Mind

Usual theoretical foundations at the heart of many endeavors in HLAI and/or computational cognitive modeling:

1

“Computer metaphor” of the mind (i.e. the concept of a computational theory of mind).

2

Church-Turing thesis.

1

Bridges gap between humans and computers:

Human mind and brain can be seen as information processing system. Reasoning and thinking corresponds to computation as formal symbol manipulation.

2

Gives account of the nature and limitations of the computational power of such a system.

Tarek R. Besold The Importance of Generality in AI

slide-13
SLIDE 13

Intelligence

No unanimously accepted definition. Scientific study of characterization and ways of measurement: Psychometrics. Working definition: Intelligence is the aggregate or global capacity of the individual to act purposefully, to think rationally, and to deal effectively with the environment.

Tarek R. Besold The Importance of Generality in AI

slide-14
SLIDE 14

Intelligence

Wide variety of psychometric tests of intelligence, ranging from tests with only one type of item (e.g., Raven’s Progressive Matrices) to varied batteries of different questions or items (e.g., Wechsler Adult Intelligence Scale). Overall range of possible items seems to cluster into groups of correlation-coupled subtests as, e.g., spatial items as opposed to verbal ones. Subtests tend to be positively correlated amongst each other and subjects scoring high on one are also fairly likely to be above average on others. ⇒ Spearman’s general factor g.

Tarek R. Besold The Importance of Generality in AI

slide-15
SLIDE 15

Psychometric Tests and Approaches in/to AI

Tarek R. Besold The Importance of Generality in AI

slide-16
SLIDE 16

Psychometrics & AI

Psychometric AI (Bringsjord & Schimanksi, 2003) Psychometric AI is the field devoted to building information-processing entities capable of at least solid performance on all established, validated tests of intelligence and mental ability, a class of tests that includes not just the rather restrictive IQ tests, but also tests of artistic and literary creativity, mechanical ability, and so on. Problem: Psychometric AI does not put any constraints on the nature

  • f the mechanisms or capacities at work in the corresponding

computational model.

⇒ Psychometric AI as it stands does not rule out the possibility of

attempts employing patchwork systems with many specialized “island solutions”.

Tarek R. Besold The Importance of Generality in AI

slide-17
SLIDE 17

Psychometrics & AI

Psychometric AI (Bringsjord & Schimanksi, 2003) Psychometric AI is the field devoted to building information-processing entities capable of at least solid performance on all established, validated tests of intelligence and mental ability, a class of tests that includes not just the rather restrictive IQ tests, but also tests of artistic and literary creativity, mechanical ability, and so on. Problem: Psychometric AI does not put any constraints on the nature

  • f the mechanisms or capacities at work in the corresponding

computational model.

⇒ Psychometric AI as it stands does not rule out the possibility of

attempts employing patchwork systems with many specialized “island solutions”.

Tarek R. Besold The Importance of Generality in AI

slide-18
SLIDE 18

Specialized AI Systems and IQ Test Problems

Current landscape of AI systems for solving IQ tests: Many specialize on one particular type of item or task in employing task-specific mechanisms or in modelling domain-specific capacities. Domain- and task-specific description languages (e.g., number sequence problems). Problem-specific representations and structure computations (e.g., matrix problems). Domain-specific heuristics or specifically trained networks etc.

Tarek R. Besold The Importance of Generality in AI

slide-19
SLIDE 19

Specialized AI Systems and IQ Test Problems

Problem: Limited value in HLAI context – little hope of generalizing and integrating these isolated solutions into a computational system (re-)creating general domain-independent human-level intelligence. Solution: Study and development of a priori domain-general mechanisms and computational models of cross-domain cognitive capacities.

Tarek R. Besold The Importance of Generality in AI

slide-20
SLIDE 20

Specialized AI Systems and IQ Test Problems

Problem: Limited value in HLAI context – little hope of generalizing and integrating these isolated solutions into a computational system (re-)creating general domain-independent human-level intelligence. Solution: Study and development of a priori domain-general mechanisms and computational models of cross-domain cognitive capacities.

Tarek R. Besold The Importance of Generality in AI

slide-21
SLIDE 21

Towards General Mechanisms and Capacities

Tarek R. Besold The Importance of Generality in AI

slide-22
SLIDE 22

General Mechanisms: IGOR2 and Program Learning

Program learning as general mechanism: IGOR2 Example driven, analytical strategy approach, integrating concepts from functional programming (e.g., pattern matching) and from ILP (possibility to use background knowledge for induction and invention of sub-programs). Hypothesis construction is based on anti-unification (i.e., least-general generalization) of sets of equations. Applied to learning correct functional programs for insertion sort, reverse, odd/even, multiplication by addition, or Fibonacci numbers.

Tarek R. Besold The Importance of Generality in AI

slide-23
SLIDE 23

General Mechanisms: IGOR2 and Program Learning

Program learning as general mechanism: IGOR2 Example driven, analytical strategy approach, integrating concepts from functional programming (e.g., pattern matching) and from ILP (possibility to use background knowledge for induction and invention of sub-programs). Hypothesis construction is based on anti-unification (i.e., least-general generalization) of sets of equations. Applied to learning correct functional programs for insertion sort, reverse, odd/even, multiplication by addition, or Fibonacci numbers.

Tarek R. Besold The Importance of Generality in AI

slide-24
SLIDE 24

IGOR2 as General Mechanism Acquiring Productive Rules

Regularity detection and generalization in IGOR2: Algorithmic approach of IGOR2 can be seen as generic mechanism for generalizing productive rules from example experiences by observing regularities and generalizing over them. “Productive rules”: A set of rules is productive when it can be applied to input of arbitrary complexity. (Chomsky, 1959) Not limited to example-based construction of recursive programs: Induction problems can be found in problem solving (e.g., Tower

  • f Hanoi), reasoning over transitive relations (e.g., ancestor), and

intelligence test problems (e.g., number series). Premise: If IGOR2 is a generic approach to induction of productive rules, then it should also be applicable to problems from such different domains without any adaptation of the algorithm.

Tarek R. Besold The Importance of Generality in AI

slide-25
SLIDE 25

IGOR2 as General Mechanism Acquiring Productive Rules

Regularity detection and generalization in IGOR2: Algorithmic approach of IGOR2 can be seen as generic mechanism for generalizing productive rules from example experiences by observing regularities and generalizing over them. “Productive rules”: A set of rules is productive when it can be applied to input of arbitrary complexity. (Chomsky, 1959) Not limited to example-based construction of recursive programs: Induction problems can be found in problem solving (e.g., Tower

  • f Hanoi), reasoning over transitive relations (e.g., ancestor), and

intelligence test problems (e.g., number series). Premise: If IGOR2 is a generic approach to induction of productive rules, then it should also be applicable to problems from such different domains without any adaptation of the algorithm.

Tarek R. Besold The Importance of Generality in AI

slide-26
SLIDE 26

IGOR2 as General Mechanism Acquiring Productive Rules

Advantageous properties of IGOR2:

1

Learning on the symbol- or knowledge-level (corresponds to typical approaches in cognitive modelling where it is assumed, that information in working memory can be inspected and verbalized):

Same representation language for examples and hypotheses (HASKELL or MAUDE programs). Semantic domain information given by the definition of algebraic data types.

2

Learning is based on hypotheses building upon regularities detected in given observations:

More “cognitively plausible” than arbitrary generate-and-test. Based on anti-unification, allowing for an elegant algorithmic approach to generalization learning.

3

No domain-specific heuristics other than a form of “Ockham’s razor”: Bias to prefer hypotheses with fewer rules and fewer recursive calls.

Tarek R. Besold The Importance of Generality in AI

slide-27
SLIDE 27

IGOR2 as General Mechanism Acquiring Productive Rules

Advantageous properties of IGOR2:

1

Learning on the symbol- or knowledge-level (corresponds to typical approaches in cognitive modelling where it is assumed, that information in working memory can be inspected and verbalized):

Same representation language for examples and hypotheses (HASKELL or MAUDE programs). Semantic domain information given by the definition of algebraic data types.

2

Learning is based on hypotheses building upon regularities detected in given observations:

More “cognitively plausible” than arbitrary generate-and-test. Based on anti-unification, allowing for an elegant algorithmic approach to generalization learning.

3

No domain-specific heuristics other than a form of “Ockham’s razor”: Bias to prefer hypotheses with fewer rules and fewer recursive calls.

Tarek R. Besold The Importance of Generality in AI

slide-28
SLIDE 28

IGOR2 as General Mechanism Acquiring Productive Rules

Advantageous properties of IGOR2:

1

Learning on the symbol- or knowledge-level (corresponds to typical approaches in cognitive modelling where it is assumed, that information in working memory can be inspected and verbalized):

Same representation language for examples and hypotheses (HASKELL or MAUDE programs). Semantic domain information given by the definition of algebraic data types.

2

Learning is based on hypotheses building upon regularities detected in given observations:

More “cognitively plausible” than arbitrary generate-and-test. Based on anti-unification, allowing for an elegant algorithmic approach to generalization learning.

3

No domain-specific heuristics other than a form of “Ockham’s razor”: Bias to prefer hypotheses with fewer rules and fewer recursive calls.

Tarek R. Besold The Importance of Generality in AI

slide-29
SLIDE 29

Application Examples of IGOR2

Problem-Solving (in cognitive psychology and AI planning):

Tower of Hanoi problems Shortest path problems Learning phrase structure grammars from example sentences Learning of recursive relations Detecting the transitivity of the is-a relation in a concept hierarchy

Intelligence tests

Number series

Tarek R. Besold The Importance of Generality in AI

slide-30
SLIDE 30

General Capacities: HDTP and Analogy

Analogy as general cognitive building block: Heuristic-Driven Theory Projection (HDTP) Generalization-based theory and model computing analogical relations and inferences between domains given in many-sorted first-order logic (FOL) representations. Domains given as finite axiomatizations, system tries to align pairs of formulae from the two domains by means of restricted higher-order anti-unification.

Tarek R. Besold The Importance of Generality in AI

slide-31
SLIDE 31

General Capacities: HDTP and Analogy

Analogy as general cognitive building block: Heuristic-Driven Theory Projection (HDTP) Generalization-based theory and model computing analogical relations and inferences between domains given in many-sorted first-order logic (FOL) representations. Domains given as finite axiomatizations, system tries to align pairs of formulae from the two domains by means of restricted higher-order anti-unification.

Tarek R. Besold The Importance of Generality in AI

slide-32
SLIDE 32

HDTP: The Rutherford analogy

Tarek R. Besold The Importance of Generality in AI

slide-33
SLIDE 33

HDTP: The Rutherford analogy

Tarek R. Besold The Importance of Generality in AI

slide-34
SLIDE 34

HDTP for Knowledge Transfer and Concept Formation

Advantageous properties of HDTP:

1

Many-sorted FOL languages are expressive and fairly general modeling languages.

2

Purely syntax-based generalization approach is necessarily domain-general.

3

Semantics can nonetheless be addressed due to sortal structure

  • f representation language and re-representation capabilities of

the system.

Tarek R. Besold The Importance of Generality in AI

slide-35
SLIDE 35

Application Examples of HDTP

Modelling classical analogy scenarios (Rutherford’s analogy, heat-flow analogy, etc.). Modelling sensorymotor-based transfer learning in mathematics and physics education and teaching. Modelling potential inductive analogy-based process for establishing mathematical concepts starting out from concrete experiences. Modelling essential parts of concept blending processes (with current focus on mathematics and music).1

1EU-FP7 FET-Open project “COINVENT” (grant no. 611553).

Tarek R. Besold The Importance of Generality in AI

slide-36
SLIDE 36

Principles and Conclusions

Tarek R. Besold The Importance of Generality in AI

slide-37
SLIDE 37

Key System Properties for Development Towards HLAI

1

General level of description: Situated on symbol level, using expressive representations and, thus, allowing for many different domains to be modelled.

2

Cross-domain mechanisms/capacities:

Technically based on a very general and domain-independent mechanism (e.g., anti-unification): Not bound to task or domain, presumably involved in many different cognitive processes (e.g., generalization, representation alignment). Modeling general capacity (e.g., analogy): Not bound to task or domain, presumably involved in many different cognitive behaviors (e.g., knowledge transfer, concept combination).

3

Possibility of semantics-sensitive computations: Possibility to take semantic aspects of domains into account without necessarily requiring a case-specific approach.

Tarek R. Besold The Importance of Generality in AI

slide-38
SLIDE 38

(Re-)Anchoring General Approaches in AI

Some suggestions for the field:

1

Dare to work on problems from HLAI, and...

2

...dare to talk about doing so.

3

Build a repository of challenge problems from different domains with relevance for HLAI, encoded in...

4

...a common description language similar to PDDL but even more general.

5

Take your approach(es) and architectur(es), put them to the test, and make the result(s) public. Evaluation criterion: An HLAI system scores higher if it can solve problems from more than one domain while still performing reasonably well on a representative sample of tasks compared to special purpose systems or human performance.

Tarek R. Besold The Importance of Generality in AI

slide-39
SLIDE 39

(Definitely Not) The End!

Thank you for your attention! Questions, comments, suggestions, criticism: Tarek R. Besold (tarek.besold@uni-osnabrueck.de) Ute Schmid (ute.schmid@uni-bamberg.de)

Tarek R. Besold The Importance of Generality in AI