A bidirectional vector associative memory architecture with application to neural networks

A bidirectional architecture for associative memory (AM) capable of vector arithmetic operations is proposed. Conventional bit-operation and word-operation are generalized to row and column operations, respectively. By introducing a pair of masking and tagging mechanisms, the proposed architecture demonstrates a symmetric functionality such that associative memory can be performed in both column and row directions. A set of built-in vector arithmetic and logic units (VALU) are designed to perform the basic vector operations, which offers O(max{n,m}) speed-up for vector operations than conventional AM at a linear (O(n+m)) cost for realizing an n/spl times/m AM array. As an applicational example, an associative processing implementation of artificial neural networks is considered.