One approach to understand classification by neural networks

This paper addresses the problem of understanding trained neural networks. We have developed methods for extracting a clear decision scheme from a neural network trained to classify. Our method is essentially constraint-based as all weights of a neural network are forced to be in the finite set of values {-1; 0; 1}. Training of the so-constrained neural network consists in adding a penalty term to the standard backpropagation error function and gradually increasing its importance. To develop and validate our method we deal with a real-world problem of the protein secondary structure prediction. Biological results were obtained using the proposed method.