Abstract

A neural network with threshold logic and feedback¹ is a plausible model for associative memory. It can be used directly to store and retrieve information with robustness and error-correction capability. It can also perform certain computations naturally, such as nearest-neighbor search. However, when a general computational problem is addressed, a special network is designed for the problem based on some energy function. The network is usually of large size, and the solution is approximate. By the efficiency of computation, we mean a quantitative measure of how cost-effective this approach is and how far it is from the theoretical limits. The efficiency of information storage in neural networks has been determined,² but the computational aspect involves essentially different ideas. The assessment of the raw computational power of a network, the complexity of a problem, and the efficiency of embedding the problem in the network are the main issues.

© 1985 Optical Society of America