Edge-Side Highly Configurable Accelerator With Efficient Automatic Structure Search Scheme

Abstract As the integration of artificial intelligence and industry becomes increasingly interwined, there is a growing focus in both academia and industry on promoting neural networks to achieve real-time high-performance implement on resource-limited devices. Edge computing usually suffers limitations in terms of storage space, size, and power consumption, it’s a great challenge to deploy exisiting convolutional neural networks (CNN) such devices. To tackle this challenge, one of the effective approachs is network pruning, which aims to reduce the size of neural networks by elimination unnecessary parameters and connections. However, most existing network pruning schemes formulate the importance criteria manually, which introduced a lot of additional parameters, leading to insufficiently stable pruning effects, and the setting of design and parameters for pruning methods heavily relies on expert experience. In this paper, we proposed a new channel pruning method based on subtractive average optimizer, which can automatically search for the optimal network model substructure, called SABOPruner. In addition, a highly configurable lightweight and high-performance CNN accelerator is proposed. On this basis, we then designed a pipelined inference architecture based on local-aware region convolution to achieve highly parallel processing of fearures and weights and buffer scheduling. Inorder to verify the effectiveness of the above approach,VGG16 and ResNet56 were deployed on Xilinx XC7Z035 platform, and the experimental results showed that VGG16 can achieve 4.4 frames per second (FPS) and 1.26 times higher memory density compared to state-of-the-art designs, and ResNet56 can achieve 40FPS based on CIFAR-10 dataset.%MCEpastebin%