Utilizing High-Dimensional Neural Networks for Pseudo-orthogonalization of Memory Patterns

Abstract

Hebbian learning rule is well known as a memory storing scheme for associative memory models on neural networks. However, this rule doesn’t work well in storing correlated memory patterns. Recently, a new method has been proposed based on pseudo-orthogonalization by XOR masking of original memory patterns with random patterns in order to overcome this problem. In this paper, we propose an extended method for pseudo-orthogonalization of memory patterns utilizing complex-valued and quaternionic neural networks. We demonstrate that Hebbian learning rule successfully stabilizes the correlated memory patterns, and these networks can retrieve the stored patterns corresponding to the external stimuli.