The Metacat Project: A Self-Watching Model of Analogy-Making

抄録

This paper presents a broad overview of the Metacat project, an extension of the Copycat computer model of fluid concepts, high-level perception, and analogy-making. Copycat models the complex, subconscious interplay between concepts and perception that gives rise to the flexible human ability to perceive apparently-dissimilar situations as being “the same”. A key feature of the architecture is the emergence of statistically-robust, high-level behavior from the interactions of many small, low-level, nondeterministic processing agents. All processing occurs through the collective actions of many agents working in parallel on different aspects of an analogy problem, without any higher-level executive process controlling the course of events. Current work on Metacat is focused on extending the Copycat model in a way that permits it to create much richer representations of the analogies it makes. This involves incorporating a long-term memory into the architecture, along with a “self-watching” ability, so that the program can recognize, remember, and recall important patterns that occur in its own processing as it solves analogy problems. Using this higher-order “meta-level” information, analogies can be compared and contrasted in an insightful way, allowing Metacat to understand and explain its answers in a way that Copycat cannot. Metacat's relation-ship to other work in AI and cognitive science is also examined, in particular work on case-based reasoning and derivational analogy.