Abstract

The notion of competition and cooperation among patterns of activity is common to many models of neural networks, for example, those employing lateral inhibition. We show how these can be implemented using a combination of saturable gain and loss in a feedback configuration. Patterns compete for the energy supplied by the gain medium, but they cooperate in saturating the loss mechanism. Feedback provides the necessary iteration so that a winner can prevail. By judicious choice of optical geometry, one can tailor the balance of the two interactions. For associative memory, small signal loss magnitude is chosen to exceed small signal gain, but the loss is made to saturate at a lower intensity. Gain is placed in an image plane while loss is placed in a Fourier plane. This gives a dynamic indicative of multistability among the patterns.

© 1988 Optical Society of America