Kohonen Feature Map Associative Memory with Area Representation

H. Abe and Y. Osana (Japan)


Kohonen Feature Map, Associative Memory, Sequential Learning, Area Representation


In this paper, we propose a Kohonen Feature Map asso ciative memory with area representation which can learn patterns successively. This model is based on the Kohonen Feature Map associative memory and the area representa tion. Most of the conventional models which can learn pat terns successively are based on the associative memories, and their storage capacities are small because their learn ing algorithm is based on the Hebbian learning. On the other hands, the Kohonen Feature Map associative mem ory based on the local representation has been proposed. It has large storage capacity, but it has not enough robust ness for damaged neurons. In the proposed model, the area representation which is an intermediate representation of the local representation and the distributed representation is introduced and the robustness for damaged neurons is improved. Moreover, the proposed model can realize auto and hetero associations and can deal with not only binary or bipolar patterns but also analog patterns. We carried out a series of computer experiments and conīŦrmed the effec tiveness of the proposed model.

Important Links:

Go Back