A Fine-grained Sparse Neural Network Accelerator for Image Classification

The ever-increasing demand sets higher standards for the computational accuracy and network complexity of neural networks, which undoubtedly brings challenges to its further development, including more complex calculations, huge amounts of data, and limitations on processor memory bandwidth, etc. The Artificial Intelligence (AI) Accelerator can accelerate the calculation of the model, which can increase the speed of the model in the inference process. We design a sparse network accelerator for image classification in this paper. The main contributions of this work are summarized as follows: (1) We designed a fixed-scale sparse convolutional neural network accelerator, which can increase the forward reasoning speed by 1.96 times on the Resnet50 network based on NVDLA. (2) A training method for this sparse network is proposed, which can achieve flops pruned of 48.2% with only 1.4% reduction of accuracy rate for Resnet50 on ILSVRC 2012.