[ CogSci Summaries home | UP | email ]

Marshall, J. B. (2002). Metacat: A self-watching model of analogy-making and perception. In W.D. Gray & C.D. Schunn (Eds.), Proceedings of the 24th Annual Conference of the Cognitive Science Society (pp. 631-636). Mahaw, NJ: Lawrence Erlbaum Associates.


author = 	{Marshall, Jim},
title =		{Metacat: A self-watching model of analogy-making and perception},
journal =	{Proceedings of the 24th Annual Conference of the Cognitive Science Society},
pages = 	{631-636},
year = 		{2002},
editor =    	{W.D. Gray and C.D. Schunn},
address=    	{Mahaw, NJ},
publisher = 	{Lawrence Erlbaum Associates} 



Author of the summary: Asaf Parush, 2007, asafparush@hotmail.com

Cite this paper for:

Metacat was modeled after Copycat (see article summary here ).

Copycat lacks elaborate self-watching, in response, Metacat was developed to include better self-watching.


-purpose of program is to model analogy making and self-awareness. -self-awareness arises from a sub-cognitive substrate which, in itself, cannot support it. Copycat "was developed as a model of the complex interplay between bottom-up and top-down processes in the mind, which together enable humans to perceive analogies between different situations in remarkably flexible ways" (pg 1).

Copycat architecture

has two main components; - slipnet: long-term memory of concepts about the letter-string world (pg 1) - Examples of Slipnet concepts: 'identity', 'predecessor' - workspace: short-term memory for perceptual structures (pg 1) - contains nondeterministic agents called codelets.

Temperature monitors state of codelets. In high temperatures few Workspace structures exist (pg 2) In low temperatures, there are more structures and thus decisions about problems gain confidence (pg 2) Temperature serves for a primitive self-watching mechanism (pg 2)

Metacat architecture

Same as Copycat but adds:

Episodic memory

- Where an answer description is stored: Workspace structures give rise to an answer: answer descriptions accumulate containing more information than the Workspace answers


- Creates themes which are "essential ideas underlying an answer" (pg 3) - "In some ways, themes are like ordinary Workspace structures. They are not initially present in the Themespace; rather, they arise during the course of a run as the result of codelet activity occuring in the Workspace" (pg 3) - Themes are compared to other answers in Episodic memory - Themes act s retrieval cues for similar answers in Episodic memory - Themes "exert strong top-down pressures on perceptual activities in the Workspace" (pg 3)

- Themes "exert strong top-down pressures on perceptual activities in the Workspace" (pg 3)

Temporal trace

- extension of Workspace - focal point of Metacat self-watching - "...stores explicit temporal record of the most important processing that occur while the program works on an analogy problem" (pg 3) - "The Trace allows Metacat to monitor subcognitive processing activity in the Workspace at a very abstract and highly-chunked level of description, enabling the program to "see" what it is doing during a run" (pg 4)

Pattern-clamping and Self-control

"Clamping a pattern alters the probability that certain types of codelets will run, or that certain types of Workspace structures will get built, effectively steering the behavior of the program in particular directions." (pg 4).

Copycat architecture is not 'self-aware' enough to clamp patterns. This lack is responsible for 'snags' in the program. A 'snag' is when the program hits a dead end and cannot proceed. 'Snags' in Copycat bring the run to a halt, whereas "If Metacat continues to hit the same snag several times in succession, a series of snag events will accumulate in the Trace, all with very similar thematic characterizations. This similarity may be noticed by codelets (the probability becoming higher as more snags accumulate), causing them to take action by clamping the "offending" themes...with strong negative activation." (pg 4)

This feature allows Metacat to monitor its own behavior using codelets like a 'Progress watcher' which determines when a pattern can be 'unclamped' and continue to run its course. If no new structures are created for some time, the run is said to hit an impasse in which case the 'Progress Watcher' advises the system to create new rules, or modify old ones, which allows the run to continue. If no new rules are found the Progress watcher codelet will remove the clamp.

The other typed of codelet responsible for monitoring the system from a higher level is the 'Jootser'.

'Jootsing' stands for Jumping Out Of The System

This happens when the system enters a feed-back loop, or repetitive 'snag' events, where no novel results may be obtained if the same procedure is employed.

By 'jootsing', the run will end with no successful results; this way the program will not cycle endlessly.

Summary author's notes: