Differentiable Soft Quantization: Bridging Full-Precision and Low-Bit Neural Networks

TL;DR

This paper introduces Differentiable Soft Quantization (DSQ), a novel method that enables stable training of low-bit neural networks by bridging the gap with full-precision models, leading to improved accuracy and deployment efficiency.

Contribution

The paper proposes DSQ, a differentiable quantization technique that automatically adapts during training to better approximate standard low-bit quantization.

Findings

DSQ outperforms existing quantization methods on various network architectures.

Training with DSQ results in higher accuracy for low-bit neural networks.

Efficient deployment of 2-4 bit DSQ on ARM devices achieves up to 1.7× speedup.

Abstract

Hardware-friendly network quantization (e.g., binary/uniform quantization) can efficiently accelerate the inference and meanwhile reduce memory consumption of the deep neural networks, which is crucial for model deployment on resource-limited devices like mobile phones. However, due to the discreteness of low-bit quantization, existing quantization methods often face the unstable training process and severe performance degradation. To address this problem, in this paper we propose Differentiable Soft Quantization (DSQ) to bridge the gap between the full-precision and low-bit networks. DSQ can automatically evolve during training to gradually approximate the standard quantization. Owing to its differentiable property, DSQ can help pursue the accurate gradients in backward propagation, and reduce the quantization loss in forward process with an appropriate clipping range. Extensive experiments over several popular network structures show that training low-bit neural networks with DSQ can consistently outperform state-of-the-art quantization methods. Besides, our first efficient implementation for deploying 2 to 4-bit DSQ on devices with ARM architecture achieves up to 1.7$\times$ speed up, compared with the open-source 8-bit high-performance inference framework NCNN. [31]