ShuffleBlock: Shuffle to Regularize Convolutional Neural Networks

Abstract

Deep neural networks have enormous representational power which has lead them to overfit on most datasets. Thus, regularizing them is important in order to reduce overfitting and to enhance their generalization capability. This paper studies the operation of channel patch shuffle as a regularization technique in deep convolutional networks. We propose a novel regularization technique called ShuffieBlock where we show that randomly shuffling small patches or blocks between channels significantly improves their performance. The patches to be shuffled are picked from the same spatial locations in the feature maps such that a patch, when transferred from one channel to another, acts as a structured noise for the later channel. The ShuffieBlock module is easy to implement and improves the performance of several baseline networks for the task of image classification on CIFAR and ImageNet datasets.