Abstract: Based on Kanerva's Sparse Distributed Memory Model (SDM), in order to recognize patterns such as Chinese character recognition and face identification problems in a large dimensional input space, an oriented-problem SDM is proposed according to such concrete cases as the frequency distribution of Chinese characters and the characteristic distributions of faces, so that the model is more practical. The Hebb learning rule in SDM is replaced by the exponential learning rule. In terms of our theoretical analysis, the signal noise ratio and memory capacity of the modified model are obviously improved and the ability of SDM is extended. The simulation of computer shows its agreement with the above analysis.

Key words: associative memory, SDM, signal-noise-ratio, memory capacity, neural networks