Dynamic Pattern Recognition Model Based on Neural Network Response to Signal Fluctuation

抄録

We have proposed a model of dynamic retrieval in associative memory based on temporal input/output correlations under a stimulus-response open scheme of neural networks. This mechanism is different from that of the conventional stationary Hopfield model in which the input signal is used only as information for the initial state of the network. Building upon the fundamental properties of the proposed model, in this paper, we newly evaluate the dependence of identification performance on the signal fluctuation level and on the number of stored patterns by introducing an accuracy rate for known (stored) and unknown (non-stored) patterns, based on the network correlation to the input signal with fluctuation. The results indicate that the dynamic scheme of network response to a fluctuating signal leads to increased efficacy and usefulness.