A method of analyzing information represented in neural networks

This paper proposes a new method of analyzing the information represented in multilayer neural networks. We define a function that estimates the degree of dependence of an output unit on an input unit. The value of this function can be used to investigate whether a learnt network detects the features in training patterns. Two series of computer simulations are presented: one for the investigation of the basic characteristic of this function and the other one for the application of our method to a simple pattern classification task.