A CONSTRAINT SATISFACTION THEORY FOR BINARY NEURAL NETWORKS

Neural networks can be viewed as open constraint satisfaction networks . According to the consideration, neural networks (NNs) have to obey a n inherent logical theory that consists of two-state decisions, weak c onstraints, rule type and strength, and identity and contradiction. Th is article presents the underlying frame of the theory that indicates that the essential reason why an NN is changing its states is the exis tence of superior contradiction inside the network, and that the proce ss by which an NN seeks a solution corresponds to eliminating the supe rior contradiction. Different from general constraint satisfaction net works, the solutions found by NNs malt contain inferior contradiction but not the superior contradiction. Accordingly, the constraints in NN s are weak or flexible. The ability of a general NN is insufficient fo r its application to constraint satisfaction problems. (C) 1996 John W iley and Sons, Inc.