Tree-structured Deep Neural Networks by Learning to Semantically Split

Abstract

We propose a novel deep neural network that is both lightweight and effectively structured for model parallelization. Our network, which we name as SplitNet, automatically learns to split the network weights into either a set or a hierarchy of multiple groups that use disjoint sets of features, by learning both the class-to-group and feature-to-group assignment matrices along with the network weights. This produces a tree-structured network that involves no connection between branched subtrees of semantically disparate class groups. SplitNet thus greatly reduces the number of parameters and required computations, and is also embarrassingly model-parallelizable at test time, since the evaluation for each subnetwork is completely independent except for the shared lower layer weights that can be duplicated over multiple processors, or assigned to a separate processor. We validate our method with two different deep network models (ResNet and AlexNet) on three datasets (CIFAR-100, ILSVRC 2012 and ImageNet-22K) for image classification, on which our method obtains networks with significantly reduced number of parameters while achieving comparable or superior accuracies over original full deep networks, and accelerated test speed with multiple GPUs. Especially, we achieved state-of-the-art accuracy on the ImageNet-22K with SplitNet based on ResNet-50. Codes available at http://vision.snu.ac.kr/projects/splitnet.