Representation Enhancement for Convolutional Neural Network Using Filter Diversity

This paper aims to improve the feature representation by diversifying CNN filters inspired by niche concept in evolution. The singular value decomposition (SVD) entropy based efficient metric for diversity is proposed In the proposed approach, filters are clustered by groups and they are calculated as differences from the center values within the groups, rather than by entire rank based comparison. This provides an effective method for increasing the substantial diversity of filters. Furthermore, the filters with low diversity are adjusted by the diversity spreading framework for better diversity in the reconstruction process. The improvement of the filter representation by performing experiments on CIFAR 10/100 data for VGG16, and ImageNet for ResNet34 is provided. Because there are no similar studies, we compare our results with respect to those of relatively relevant pruning methods in terms of classification performance accuracy as well as the pruned rates and flops. Copyright © The Korean Institute of Electrical Engineers.