Coarse Pruning of Convolutional Neural Networks with Random Masks

The learning capability of a neural network improves with increasing depth at higher computational costs. Wider layers with dense kernel connectivity patterns further increase this cost and may hinder real-time inference. We propose feature map and kernel pruning for reducing the computational complexity of a deep convolutional neural network. Due to coarse nature, these pruning granularities can be exploited by GPUs and VLSI based implementations. Further, we propose a simple strategy to choose the least adversarial pruning masks. The proposed approach is generic and can select good pruning masks for feature map, kernel and intra-kernel pruning. The pruning masks are generated randomly, and the best performing one is selected using the validation set. The sufficient number of random pruning masks to try depends on the pruning ratio, and is less than 100 when 40% complexity reduction is needed. Once the least adversarial pruning mask is selected, we prune and retrain the network in one-shot. The proposed approach therefore consumes less time compared to iterative pruning. We have extensively evaluated the proposed approach with the CIFAR-100, CIFAR-10, SVHN, and MNIST datasets. Experiments show that 60-70% sparsity can be induced in the convolution layers with less than 1% increase in the misclassification rate of the baseline network.