Context-dependent Sequential Recall by a Trajectory Attractor Network with Selective Desensitization

Masahiko Morita, Kazuhiko MURATA and Shigemitsu Morokami

Abstract: We present a model composed of nonmonotonic neurons that recalls various target patterns from the same cue pattern depending on a given context pattern, which was essentially difficult for conventional neural network models. In this model, the state of the network is projected onto a subspace by desensitizing a part of neurons depending on the context, and shifts along a trajectory attractor in the subspace to reach the target state. The model can simulate an arbitrary finite state automaton without limitations of size, keeping the merits of distributed representation.


Reprint [ PDF (317KB) ] is available