Maximum/minimum detection by a module-based neural network with redundant architecture
K. Tsutsumi, K. Nakajima · 2003
A neural network based on relaxation dynamics is known to function as a maximum/minimum detector. In such a network, it is necessary to design an adequate energy function to be minimized for the derivation of network dynamics. However, even if the feedback connections are well-tuned, the detection greatly depends on the initial states of neural cells. For example, under the condition that all or some of the maximal/minimal values in a task are the same, certain cell states may not change on a saddle point in the network dynamics. We consider the simplest case of a 2-value minimum defection task. We show how a module-based neural network with "redundant" architecture is used in an attempt to solve such problems as the initial-value dependence and deadlocking.