Bit error probability of an associative memory with many-to-many correspondence

Abstract

In associative memories (AMs) an important characteristic is the type of information that is stored in the connection weight. Conventional AM models store a one-to-one correspondence between memory patterns. In these models therefore, the cue pattern used to recall a pattern is restricted to its associated pair and noisy versions of this associated pair.

To overcome this restriction, a new model is proposed in which a many-to-many correspondence between key patterns and content patterns is stored. A content pattern can be recalled from various noisy key patterns which are sequentially provided to the network. The combination of key patterns gives variety and flexibility to the AM and increases its usefulness. In this paper, the bit error probability of the output of erroneous signal by each neuron is estimated. Computer simulations show good agreement with analytical results.