Research On Associative Memory For One To Many Associations Based On Incidence Of Patterns

The realization of associative memory is always the one of researching directions of artificial neural networks and the key problem is to realize associative memory for many to many associations. The core of associative memory for many to many associations is how to realize associative memory for one to many associations, that is, how to recognize common items in memory patterns. Aiming at this key problem, based on others' researches, this thesis proposed a new model of associative memory for one to many associations based on incidence of patterns and discussed associative memory properties of new model for realizing associative memory for one to many associations. Main contributions of this thesis are summarized as following three aspects:I To construct a new associative memory for one to many associations based on incidence of patterns. In this thesis, the properties of associative memory are analyzed firstly for one to many and many to many models proposed before; and then the incidences of locations are introduced among items of storing patterns to classic discrete Kosko bipolar neural networks and to realize associative memory for one to many using multi modules of relative simple BAM.II To guarantee exact retrieves of memory patterns stored in new model in theory. For inner limitations of associative memory for one to many models and many to many models proposed before, all of them can not guarantee exact retrieves of memory patterns stored in themselves in theory. The thesis has introduced the incidences of items of memory patterns to new model so that it can guarantee exact retrieves and avoid the continuity limitation in some sense of memory patterns faced with discrete Kosko bipolar neural networks.III To analyze knowledge-increasing ability of new model. Dynamic adjustment of neural network structure and knowledge-increasing learning is a new field of research in neural networks. Classic discrete bipolar neural networks have no abilities to learn knowledge increasingly. However, associative memory model for one to many associations based on incidence of patterns proposed in this thesis, through adding new neurons to neural networks of this model, does no forget information stored before as well as has increased memory information and an ability of knowledge increasing.Finally, the research of this thesis make experiments for realizing of associative memory for one to many and many to many associations based on classic neural networks and is good to engineering practices of associative memory based on models with simple structure for one to many associations.