Systematic Model Complexity Reduction by Elimination of Irrelevant Layers in Convolutional Neural Networks

Abstract

Neural networks were treated as black boxes for a long time. Previous works have unearthed what aspects of an image were important for convolutional layers at different positions in the network. This was done using deconvolutional networks. In this paper, we examine how well a convolutional neural network performs when those convolutional layers which are relatively unimportant for a particular image (i.e., the image does not produce one of the strongest activations) are skipped in the training, validating, and testing process.