FLQ: Design and implementation of hybrid multi-base full logarithmic quantization neural network acceleration architecture based on FPGA

Longlong Zhang, Xiang Hu, Xuan Liao, Tong Zhou, Yuanxi Peng

Abstract: As deep neural network (DNN) models become more accurate, problems such as large model parameters and high computational complexity have become increasingly prominent, leading to a bottleneck in deploying them on resource-limited embedded platforms. In recent years, logarithm-based quantization techniques have shown great potential in reducing the inference cost of neural networks. However, current single-model log-quantization has reached an upper limit of classification performance, and little work has investigated hardware implementation of neural network quantization. In this paper, we propose a full logarithmic quantization (FLQ) mechanism that quantizes both weights and activation values into the logarithmic domain, compressing the parameters on AlexNet and VGG16 model by >6.4 times while maintaining an accuracy loss of within 2.5 % compared with benchmarking. Furthermore, we propose two optimization solutions for FLQ: activation segmented full logarithmic quantization (ASFLQ) and multi-ratio activation segmented full logarithmic quantization (Multi-ASFLQ), which can better balance the numerical representation range and quantization step. Under the condition of weight quantization of 5 bits and activation value quantization of 4 bits, the optimization methods proposed in this paper can improve the TOP1 of the VGG16 network model by 1 % and 1.6 %, respectively. Subsequently, we propose an implementation scheme of computing unit corresponding to the optimized FLQ mechanism above, which can not only convert multiplication operations into a shift operation but also integrate functions such as different ratio logarithmic bases and sparsity processing for activation, minimizing resource consumption as well as avoiding unnecessary calculations. Finally, we experiment with VGG19, Retnet50, and Densenet169 models, proving that the proposed method can achieve good performance under lower bit quantization. © 2001 Elsevier Science. All rights reserved