Abstract
High-order correlation matrix associative memory (HCMM) is proposed and examined in this paper.
The conventional correlation matrix associative memory (CMM) so far studied utilizes only the second-order correlation involved in the signals and is not concerned with the third and higher order correlations which might be quite effective for association if utilized. This causes the major draw-back of the CMM, that is, the accuracy of the recall decreases considerably when the signals are linearly dependent. Thus, the memory capasity of the CMM is Iimited not by the number of the memory cells but by the dimension of the signals. This is the great incovenience in the practical use of the associative memory.
The HCMM is introduced to dissolve the drawback of the CMM. Examinations of the HCMM's principle, the correct recall rate to random signals, the memory efficiency and the robustness of the memory for a partial break-down make clear the following preferable properties of the HCMM:
As the order of correlation increases;
- the accuracy, of recall increases when the signals are linearly dependent,
- the memory capacity, increases while keeping the same accuracy of recall.
- the memory capacity per cell decreases,
- the robustness of the memory for a partialb reak-down increases.
