Multimodule associative memory for many-to-many associations

Abstract

In this paper, we propose a novel associative memory model called MultiModule Associative memory for Many-to-Many Associations, (MMA) 2 for short. The proposed (MMA) 2 consists of multiple modules and each module has a Hopfield type of associative memory. In the (MMA) 2, a memory item is regarded as several divided patterns. Each pattern is assigned to each module and the patterns are related to each other in the learning. Unlike a single-layered conventional associative memory, the (MMA) 2 can recall a complete memory item from even a single part of it owing to the multiple modules structure. Even if a part of a memory item that is common to the other memory items is given to the proposed (MMA) 2, all items that relate to the input can be recalled: that is, the proposed (MMA) 2 can deal with the set of memory items which includes one-to-many relations and many-to-many relations such as (A 1,B 1,C 1, …), (A 1,B 2,C 2,…), (A 2,B 2,C 3,…), … . In order to memorize and recall such very complicated training data, the (MMA) 2 employs pseudo-noise (PN) patterns, transformation of distributed patterns into locally represented patterns and the logical operations. These techniques contribute to avoid producing a mixed unknown pattern, which consists of a superimposed pattern of some stored patterns and the cross-talk noise and interferes with recalling correct patterns. A number of computer simulation results show the effectiveness of the proposed (MMA) 2. Furthermore, we show that the (MMA) 2 can deal with a knowledge processing.