A Kernel-level Group-sparsity Pruning Scheme for Efficient ConvNets

Deep learning models, particularly convolutional neural networks (i.e. CNNs or ConvNets), perform admirably in many applications. Network pruning (which includes both structured and non-structured pruning) can minimize run time and the number of parameters of the CNNs. However, structured pruning always prunes some key units incorrectly, while non-structured pruning cannot easily reduce computation in real-world applications. In this paper, we present a kernel-level pruning scheme based on the Winograd minimal filtering algorithm. While minimizing the number of parameters, it can retain excellent accuracy and computation efficiency. We use a mask matrix in the Winograd domain to skip the trimmed kernel, which results in fewer matrix multiplication operations in practice. Experiments on the ILSVRC'15 dataset reveal that the accuracy of our method is 7.97 times higher compared to the filter-level pruning method on VGG-16 when the pruning ratio is 90%.