Neural Network Pruning Using Discriminative Information for Emotion Recognition.

In the last years, the effort devoted by the scientific community to develop better emotion recognition systems has been increased, mainly impulsed by the potential applications. The Boltzmann restricted machines (RBM) and the deep machines of Boltzmann (DBM) are models that, in recent years, have received much attention due to their good performance for different issues. However, it is usually difficult to measure their predictive capacity and, specifically, the individual importance of hidden units. In this work, some measures are computed in the hidden units in order to rank their discriminative ability among multiple classes. Then, this information is used to prune those units that seem less relevant. The results show a significant decrease in the number of units used in the classification at the same time that the error rate is improved.