A neural network model for extracting correlated information in sensory integration

In our previous work, we have proposed a neural network model for integrating different kinds of sensory information. After the network learning was completely performed, the internal representation of an object was formed in the third layer in which different kinds of information were integrated. However, the information about object's properties could not be extracted explicitly in the acquired internal representation. In this study, we modify the former model to extract object's properties. The third layer in this model is divided into a subset of correlation unit and subsets of no-correlation unit. After sufficient learning, the neuron activity in the correlation unit changes monotonically with a correlated parameter. This result suggests that the proposed model can extract correlated parameters and have a good ability of generalization.