Block-sparse CNN: towards a fast and memory-efficient framework for convolutional neural networks

Abstract

In this paper, we propose a block-sparse convolutional neural network (BSCNN) architecture that converts a dense convolution kernel into a sparse one. Traditional convolutional neural networks (CNNs) face the problem that an increase in the number of network parameters will lead to more model and floating-point computations, and a higher risk of network overfitting. The block-sparse convolution uses sparse factor pairs to randomize a sparse convolution kernel, which can introduce mixed information and thereby enabling the extraction of more diverse features. In the meantime, a SUMMA-based parallel computing method is adopted to achieve a lightweight storage and a fast calculation of the convolution kernel. Experimental results show that, compared with current sparse networks, the proposed framework achieves better prediction accuracy than the classical backbone networks in terms of faster floating-point operation and less storage space requirements.