Multiplying Elimination of 8-bit Low-Precision Neural Networks Exploiting Weight and Activation Repetition

摘要

Convolutiona1 neural networks (CNNs) have been applied to various applications, such as image recognition and speech recognition, and have even achieved higher prediction accuracy than human eyes. However, the computation complexity of CNN increases rapidly with the increase of network scale, and a huge number of multiplying and accumulating operations may involve for a moderate CNN instance. This may significantly prolong the training and inference process and incur large energy consumption. Thus, a few low-precision CNN acceleration methods are presented to reduce the time complexity with the price of reduced computational accuracy. But they could not essentially reduce the number of calculations. In this backdrop, this paper proposes a table lookup-based multiplication elimination method for low-precision CNNs exploiting their weight and activation repetition. In our method, a table is established in advance to store all the possible multiplying results, and a simple table lookup operation is triggered every time a multiply calculation is encountered. Analysis results have shown that our proposal can greatly reduce the computational time, memory requirement, and energy consumption of low-precision CNNs.