Abstract

An associative memory is essentially a mechanism for storing pairs of information patterns $\overline{\text{u}}$ and $\overline{\text{v}}$, so that at a later time presentation of one pattern, $\overline{\text{u}}$, will result in recall of the other, $\overline{\text{v}}$. These information patterns are represented here as time-varying vectors, $\overline{\text{u}}$,$\overline{\text{v}}$, of n discrete binary or analogue elements. These vectors could encode, for example, pixel intensities from a $\sqrt{\text{n}}\times \sqrt{\text{n}}$ 2-D image, strings of characters from a textual database, speech spectral samples, the state variables of a control system, features extracted from a scene, or the output of a robot's sensors. A large number of paradigms have been advanced for implementing associative memory. These^1,2 range from purely deterministic matrix multipliers and content-addressable digital memories to holography and randomly interconnected neural network models. The various associative memory schemes offer differing performance capabilities in terms of such measures as^1,2 information capacity, the ability to recall $\overline{\text{v}}$ when presented with only a similar (e.g., partial or distorted) version of the key pattern $\overline{\text{u}}$, discrimination ability between significantly different u's, information pattern crosstalk, hardware fault-tolerance, pretaught versus adaptive self-organizing capabilties, implementational complexity, and retentiveness/forgeting of old associations. It is sometimes desirable to distinguish an autoassociative memory, where $\overline{\text{u}}$ recalls itself, from a heteroassociative memory, where $\overline{\text{u}}$ and $\overline{\text{v}}$ differ.

© 1985 Optical Society of America