Feature Correlation Loss in Convolutional Neural Networks for Image Classification

Abstract

Feature maps in Convolutional neural networks are extracted automatically with some initialization methods and training strategies, which greatly economizes the cost of feature engineering. However, correlation between feature maps are not considered in common networks, resulting in the increase of redundant feature maps with the networks becoming more complicated. In this work, we proposed the correlation layer and designed the correlation loss, which can compute the correlation coefficient matrix of the feature maps in the last convolutional layer and optimize the weights distribution respectively. In the training phase, 2 strategies, namely the supervision and initialization are studied with Gaussian and He initialization methods for the baseline. The experimental results on CIFAR-10 dataset demonstrated that the supervision strategy for the multi-task training could efficiently reduce the correlation between the feature maps learned and increase the classification accuracy from 0.39% to 1.14% on the test set.