Phil/Psych 446, Cognitive Modeling, Week 7

7a, Cognitive Wheels in Symbolic Artificial Intelligence

Comparison: Digital and natural intelligence

	Digital intelligence	Human mind
Architecture	serial	massively parallel
Component speed	very fast	slow (1 ms)
Search processes	deep	shallow (pattern matching)
Representation	verbal, mathematical	multimodal, e.g. visual
Embodiment	simple robots	evolved
Emotional	no	yes
Accuracy	exact	approximations

Cognitive wheels

A cognitive wheel is a system that is computationally powerful but is not psychologically and biologically natural (Dennett, 1984).

It's ok for Artificial Intelligence to use cognitive wheels for technological purposes, but cognitive modelling can't, since the aim is to understand human minds.

Discussion question: what are your best candidates for cognitive wheels?

What is the opposite of a cognitive wheel? Perhaps a cognitive leg, a system that is biologically natural but not so easy to turn into technology.

Cognitive wheels in AI

Problem solving by domain-independent search

• Deep Thought is the world chess champion thanks to deep search.
• In contrast, humans do not search very deep, but instead use learned patterns, which may be general schemas or particular cases.

Logical, deductive techniques such as theorem proving

• But psychologists such as Johnson-Laird have argued that people do logical reasoning differently, by mental models.

Probability theory and Bayesian networks

• Very influential in current AI, e.g. in the Russell and Norvig text.
• But probability theory is historically recent (I. Hacking, The Emergence of Probability) and psychologically difficult (Kahneman and Tversky).
• However, it is possible that human thinking uses other means (e.g. neural networks) to approximate Bayesian reasoning.

Various learning methods (see 7b below).

Of course, it could turn out that what I think is a cognitive wheel is really what the brain does.

Cognitive legs

A cognitive leg is a biologically natural system that is hard to computerize.

Examples:

Non-verbal processes, e.g. vision

Language processing that can deal with ambiguity

Expertise arising from pattern matching

• E.g. chess, Wayne Gretsky: Skate to where the puck is going, not to where it's been.
• Combination of schemas and analogies, often non-verbal

Emotional cognitive processing

Consciousness

7b AI Learning mechanisms

Why learning is important to cognitive modelling

People learn new concepts, rules, and tasks.

People incrementally improve their ability to solve problems.

People aren't easily programmed - they have to learn to do things themselves.

We have already discussed learning of rules, frames and analogs (weeks 4-6).

Discuss: What have you learned so far in this class?

• New concepts, e.g. forward chaining, inheritance.
• New rules, e.g. schemas are powerful but inflexible.
• New skills, e.g. how to write a LISP program.
• New patterns, e.g. LISP programming types.

Learning from positive and negative examples (Winston, ch. 16)

See Winston, p. 350, for examples of arches and non-arches.

To learn the concept of an arch, it is helpful to have both positive examples and near misses that indicate what is crucial in the positive examples.

Structure: semantic network, with objects represented by nodes, and relations represented by links.

Procedure: Construct an initial description from positive examples. Evolve a better model using near misses that indicate what features are important, e.g. that the top of an arch is supported, and which features must be avoided, e.g. that the supports must not touch.

Cognitive evaluation

• Human representation of arches is primarily visual, not verbal.
• Human concepts are prototypes, not strick definitions.
• Human learning may blend examples rather than using near-miss procedures.

Learning by explaining experience (ch. 17)

Explanation-based learning systems improve the performance of a problem solver by examining how it solved previous problems and modifying it to solve future problems more efficiently. The modification is based on producing an explanation of what previously worked using the causal information in the previous solution.

E.g. learning what a cup is by explaining what makes a cup a cup. Learning what jealousy is by explaining what happened in a play.

Structure: semantic net or frame system with links or slots for causal relations.

Procedure: To understand a target example, find a similar source example with known causal structure (analog retrieval). Map the causal structure of the source to the target. Form general rules about the causal relations in the target.

Cognitive evaluation

• Requires very rich causal information
• Computationally useful, but do people ever do anything this complicated?

Learning by correcting mistakes (Winston, ch. 18)

Combines learning from examples and explanation-based learning to use failures to correct identification models.

Learning from examples using version spaces (Winston, ch. 20)

Structure: A version space keeps track of the useful information supplied by a sequence of learning examples. It has a specialization tree and a generalization tree with nodes that models that match examples.

Procedure: specialize the general models and generalized the specific models until you get a correct model that matches all observed positive examples and does not match any negative examples.

Cognitive evaluation:

• Learns only strictly defined concepts, not prototypes.
• Algorithm is elegant, but keeps track of more information than people can.

Learning by building identification trees (Winston, ch. 21)

Very widely used in machine learning and data mining

Example: given examples of UW students, construct a decision tree to classify them into artsies, mathies, etc.

Structure: a decision tree in which each node is connected to a set of possible answers, each non-leaf node is conntected to a test, and each branch carries a particular test result's subset to another node.

Procedure: given examples, construct the smallest identification tree that is consistent with them.

An identification tree can be converted into a set of rules.

Cognitive evaluation

• Human concepts may be more loose and flexible than identification trees capture.
• But Ling argues that identification trees are as good as neural networks at modelling human learning.

Version spaces look like cognitive wheels, but decision trees may not be.

Assignment 4, due March 12.

Don't forget that the Project Plan is due March 5.

Note: all assignments must be handed in on paper in class or to HH 365.

Updated Feb. 12, 2002

Back to Phil/Psych 446