Any recognition process involves making a connection between a concept representation stored in the system’s memory and a stream of observational data. Present recognition systems attempt to recognize objects by matching descriptions with the data stream. If the input data satisfies rules characterizing an object, the object is recognized. To implement such a system for practical tasks, a very large number of rules may be required.
This aspect severely limits present recognition systems, as it prevents them from being applied to the recognition of a large number of objects. In contrast to this approach, humans can recognize objects from a great variety of different cues, without “matching” rules. For example, they can recognize a known person from seeing a face, a silhouette, the characteristic way of walking, hearing the person’s voice, or even from observing the person’s gesticulation or seeing his/her shoes.
The dynamic recognition (DR) approach (initially proposed by Michalski in 1986) overcomes this problem by using inductive inference to dynamically determine discriminant object descriptions from characteristic object descriptions, and this allows the system to avoid matching rules. Only one characteristic description per concept is stored in memory. Potentially, the DR method can efficiently handle a great variety of different practical recognition problems. An initial implementation of the system has strongly supported the theoretical expectations.
Michalski, R, S,, “Dynamic Recognition: An Outline of Theory of How to Recognize Concepts without Matching Rules,” Reports of the Intelligent Systems Group, ISG 86-16, UIUCDCS-F-86-965, Urbana, 1986.
Michalski, R. S., “A Variable-Valued Logic System as Applied to Picture Description and Recognition,” Chapter in the book, Graphic Languages, F. Nake and A. Rosenfeld (Editors), North-Holland Publishing Co., 1972.
For more references, see publications section.