On the problem of self-organization in neurocomputing

The results are presented of a theoretical analysis of self-organizing (SO) natural objects with such common features as learning, memorizing (acquisition), and recognition. The knowledge of the mechanisms of learning and of the genesis of SO objects had made it possible to formulate general principles for creating artificial counterparts capable of learning and recognition. Some of these principles have been implemented in a software model of a learning visual pattern recognizer. This model is capable of learning and recognizing gray-level and/or colored images on a screen of 64 by 64 pixels; recognition results are not influenced by changes in the background intensity and color (which may be inverse), and the images may consist either of line patterns or inked surfaces. Recognition may be performed in real time and is invariant to rotation, displacement, and 1:7 zooming within the boundaries of the screen. The signal-to-noise ratio is satisfactory for practical purposes. This recognizer detects a pattern which resembles most of those it has already memorized, even if it is a part of another pattern or contains other patterns. The recognizer can memorize about a hundred patterns