Characteristics of the associative memory trained by the learning-unlearning algorithm

Characteristics of the associative memory model trained by the learning-unlearning algorithm are described. The unlearning method was investigated in order to increase the stability of the specific patterns and decrease the probability of the finding of spurious patterns. However, the characteristics of the model trained by the algorithm including the method are not clearly related to the performance of the model, that is, the capacity of the memory, the probability of finding memorized patterns, or the separability of memorized patterns. The outer product algorithm for training a model is familiar and easy, but the performances of the model trained with this algorithm are insufficient in practical use. In order to improve the performances, the learning-unlearning algorithm which uses the unlearning method is proposed. Through a formal description of the algorithm, learnability and restrictions of the model are described. The accessing probability of the model trained by the proposed algorithm shows a much higher value than that of the model trained with the other product because this algorithm can delete many spurious states