Self-Creating and Organizing Neural Networks with Weight Duplication

We propose a new self-creating and self-organizing neural networks utilizing weight duplication. It is well known that columnar structures in the brain play a great role in visual information processing of early vision. In a columnar network, there are many cells, which represent the similar features. These characteristics are important for the robust information processing. The proposed model consists of competitive and input layers. Competitive learning modifies weights of the activated node. A new node is created depending on the activation frequency. The weights of a daughter node are duplicated from the mother node. The mother node has refractory period just after the creation. The weights of the daughter node and those of her mother node will be similar after the refractory period. A mother-daughter relationship will represent the hierarchical structure of input data. It is possible to organize the column by itself with similar features, which has hierarchical structure. The represented features will become precise when descending the hierarch y. Some simulations are carried out to show the feasibility of the proposed model.