ABSTRACT
In this paper, we present an approach to understanding trained neural networks. We start with a simple observation that a neural net trained on a classification task and having weights clustered around – 1,0 and 1 can be interpreted as a majority-vote rule. A weight penalty technique is used to force the weights to discrete values. This technique requires quite a bit of tuning of control parameters over time. A large number of parameters involved makes our method difficult to replicate and its generality doubtful. To remede such a situation we propose a generalization of this empirical method through a Bayesian framework and Gaussian mixture models. We conduct our experiments on the real-world problem of protein secondary structure prediction. With our approach we have extracted a majority vote rule explaining the neural net’s classification of protein secondary structure without losing prediction accuracy.
