Abstract

Associative memory based on the Hopfield neural network model can be used directly to store and retrieve information with robustness and error-correction capability. Recently an attentive associative memory was reported which provides the flexibility of rapidly and arbitrarily changing the strengths of the stored images. With slight modification of the conventional retrieving formula we propose a new mechanism for incorporating another kind of attention, i.e., attention-to-bits or the relative significance of each bit in the input images. This relative significance is implemented by multiplying the proper weights on each bit of the input images and can be altered arbitrarily. Computer simulation shows that this associative memory indeed has a tendency to retrieve the stored images with minimum weighted-Hamming distances defined as sums of the weights for all the incorrect bits. When only partial input images are available or parts of the images are of critical importance, this attentive-bits associative memory may be of great use.

© 1987 Optical Society of America