Test-cost-sensitive Convolutional Neural Networks with Expert Branches

Abstract

It has been proven that deeper convolutional neural networks (CNN) can result in better accuracy in many problems, but this accuracy comes with a high computational cost. Also, input instances have not the same difficulty. As a solution for accuracy vs. computational cost dilemma, we introduce a new test-cost-sensitive method for convolutional neural networks. This method trains a CNN with a set of auxiliary outputs and expert branches in some middle layers of the network. The expert branches decide to use a shallower part of the network or going deeper to the end, based on the difficulty of input instance. The expert branches learn to determine: is the current network prediction is wrong and if the given instance passed to deeper layers of the network it will generate right output; If not, then the expert branches stop the computation process. The experimental results on standard dataset CIFAR-10 show that the proposed method can train models with lower test-cost and competitive accuracy in comparison with basic models.