On reducing the influence of noise in a new model for optimal linear associative memory

Abstract

We propose a new model for linear associative memory. In this work, a synaptic matrix composes of not only the stored input and output patterns, but also of injected attached patterns (constructed by designers) which are weighted periodic inverse-repeat pseudo random patterns.When the injected patterns are taken to be the stored input patterns, one obtains Kohonen's model. As such, Koh* nen's model is a special case of the model proposed here. Furthermore, when the real pattern is contaminated by (different kinds of) color noise, recalling the stored pattern is superior than that obtained from Kohonen's pseudo-inverse learning rule: our learning rule can reduce the color noise influence on the optimal linear associative memory.We believe that injecting attached patterns to neurons is a new vista in neural network theory. The learning rule proposed here is shown to be optimal in the least mean square sense. We illustrate the theoretical results with specific computer simulation.the Department of Mathematics, Harvard University for providing the initial graphical data. Prof Alan Yuille from the Department of Computer Science and John Yao from the Division of Applied Sciences, Harvard University, have improved greatly this manuscript with many helpful advises. The algorithms for this model were implemented by Dr Bainan Li in a SUN work station in the computer laboratories of the