Neural Network to Control Output of Hidden Node According to Input Patterns

In an ordinary artificial neural network, individual neurons have no special relation with an input pattern. However, some knowledge about how the brain works suggests that an advanced neural network model has a structure in which an input pattern and a specific node correspond, and have learning ability. This paper presents a neural network model to control the output of a hidden node according to input patterns. The proposed model includes two parts: a main part and control part. The main part is a three-layered feedforward neural network, but each hidden node includes a signal from the control part, controlling its firing strength. The control part consists of a self-organizing map (SOM) network with outputs associated with the hidden nodes of the main part. Trained with unsupervised learning, the SOM control part extracts structural features of input space and controls the firing strength of hidden nodes in the main part. The proposed model realizes a structure in which an input pattern and a specific node correspond, and undergo learning. Numerical simulations demonstrate that the proposed model has superior performance to that of an ordinary neural network.