SONNET: a self-organizing neural network that classifies multiple patterns simultaneously

The fundamentals are presented of a self-organizing neural network (SONNET) that can classify multiple distinct patterns simultaneously. The network consists of two fields, F⁽¹⁾ and F ⁽²⁾. Patterns are registered at F⁽¹⁾ and classified at F⁽²⁾. The spatial patterns at F⁽¹⁾ continually evolve; therefore, learning must be done in realtime. F⁽²⁾ is an on-center off-surround network that obeys winner-take-all dynamics. At F⁽²⁾, new classifications can form without degrading previous classifications; therefore, the learning is stable. F ⁽²⁾ is not a homogeneous field. Nodes learn different output characteristics so that different nodes can respond to different size patterns. Nonhomogeneous inhibitory connections form at F⁽²⁾ so that nodes compete only with other nodes coding similar patterns. This allows multiple F⁽²⁾ nodes (each representing a distinct pattern) to activate simultaneously