Analysis of a correlation-type associative memory with one-to-many associations

The associative memory model HASP, that consists of a heteroassociative network and an autoassociative one. call associate a key item with more than one associative items. In order to analyze the recalling: process in the scenario of one-to-many associations, the statistical correlation between the associative vectors and network state of the autoassociative part must be taken into account. By using a simplified model, which qualitatively captures characteristics of the HASP and also consists of a heteroassociative network and an autoassociative one. we analyzed its recalling process. By introducing an external input that is statistically correlated with a target associative vector, which is one of recalling associative vectors, the symmetry between these associative vectors can be broken. It is shown that (a) the recall performance is better when the external input is injected into the autoassociative network than when it is injected into the heteroassociative network, and (b) there is a critical correlation between the external input and the target: associative vector above which the recalling process is successed.