A Proposal of Self-Creating Type Self-Organizing Neural Networks

Abstract

A new self-creating type self-organizing neural network is proposed. It is well known that the columnar structure in the human brain plays an important role in visual information processing. In the columnar network, cells which represent similar features are gathered nearby. This structure is useful for robust in-formation processing. Realization of the structure in the computational model is effective for intelligent information processing. The key concept of the proposed model lies in the self-creation of new nodes with weight duplication. The daughter node is created based on a self-organizing neural network. The mother node has refractory period just after the creation. The weights of the daughter node and those of her mother node become similar after the refractory period. A hierarchical learning leads the mother-daughter relation-ship to represent the similar features of the columnar network. This results in self-creation of the nodes which represent similar features without dead nodes. Simulations are carried out to show feasibility of the proposed model.