A study on with the bottom-up and the top-down procedure in hierarchical neural networks

This paper presents a new framework for hierarchical neural networks. It is based on the interaction between bottom-up and topdown procedure to realize human-like cognitive processes. The proposed system consists of 3 layers, input layer, feature extraction layer and symbol layer. Input layer is a kind of associative memory and associate by itself. When the association fails, feature extraction through bottom-up procedure and attention to the parts of input pattern through top-down procedure make it possible to associate target pattern. Some simulation results show the feasibility of proposed system.