ADAPTIVE PERCEPTUAL PATTERN-RECOGNITION BY SELF-ORGANIZING NEURAL NETWORKS - CONTEXT, UNCERTAINTY, MULTIPLICITY, AND SCALE

A new context-sensitive neural network, called an EXIN (excitatory + i nhibitory) network, is described. EXIN networks self-organize in compl ex perceptual environments, in the presence of multiple superimposed p atterns, multiple scales, and uncertainty, The networks use a new inhi bitory learning rule, in addition to an excitatory learning rule, to a llow superposition of multiple simultaneous neural activations ( multi ple winners), under strictly regulated circumstances, instead of forci ng winner-take-all pattern classifications. The multiple activations r epresent uncertainty or multiplicity in perception and pattern recogni tion. Perceptual scission (breaking of linkages) between independent c ategory groupings thus arises and allows effective global context-sens itive segmentation, constraint satisfaction and exclusive credit attri bution. A Weber Law neuron growth rule lets the network learn and clas sify input patterns despite variations in their spatial scale, Applica tions of the new techniques include segmentation of superimposed audit ory or biosonar signals, segmentation of visual regions, and represent ation of visual transparency.