A New Geometry with Cross-Coupling of ART Networks

摘要

This paper demonstrates a new geometrical arrangement of adaptive resonance theory based network. Using method of minimal anatomies a neural network was constructed in an attempt to compare patterns. The anatomy incorporates two sub-networks coupled by feedback signals and an additional motor layer whose outputs reflect relationship or non-relationship among the compared patterns. Simulation results illustrates the network behaviors as emergent properties. The network with unsupervised learning is capable of generating self-defining features.