Minimizing Computation in Binarized Neural Network Inference using Partial-Filter Sharing

Binarized Neural Network (BNN), which is a variant of Convolutional Neural Network (CNN) with binary weights and binary outputs on a neuron, has emerged as a promising approach to deploy artificial intelligence on resource-restricted devices in recent years. Due to the binarized weights, there are no longer required multipliers for computation, and there exist relatively high similarities among filters as well. In this work, we propose a partial-filter sharing approach and integrate it with the state-of-the-art to reduce the hardware cost and the synthesis time onto Field Programmable Gate Arrays (FPGAs). As compared to the state-of-the-art, the LUT reduction ratio by our approach is 47.71% on average without any accuracy loss, and 62.5% synthesis time in the Tiny ImageNet layers can be also saved on average.

I. Introduction

In recent years, Convolutional Neural Networks (CNNs) have achieved a huge success in artificial intelligence. It conquered plenty of problems that were hard to deal with using computers in the early days. Currently, deploying AI on edge devices such as smartphones, wearable devices, and IOT devices has become a trend to meet people's daily need. However, a great computational challenge still remains for achieving this goal. CNNs usually contain vast numbers of floating-point parameters and require enormous numbers of floating-point operations in both training and inference phases. This scenario makes CNNs difficult to be implemented on edge devices. As a result, minimizing the computational complexity of neural network inference on edge devices has become crucial.