Automatic Network Adaptation for Ultra-Low Uniform-Precision Quantization

TL;DR

This paper introduces a neural architecture search method that adaptively expands network channels to improve ultra-low precision quantization accuracy, addressing layer sensitivity differences.

Contribution

It proposes neural channel expansion, a novel approach to optimize network structure for 2-bit quantization, enhancing accuracy while respecting hardware constraints.

Findings

Achieves state-of-the-art 2-bit ResNet50 accuracy on CIFAR10 and ImageNet.

Improves inference accuracy with fewer FLOPs and parameters.

Effectively adapts network channels to mitigate quantization errors.

Abstract

Uniform-precision neural network quantization has gained popularity since it simplifies densely packed arithmetic unit for high computing capability. However, it ignores heterogeneous sensitivity to the impact of quantization errors across the layers, resulting in sub-optimal inference accuracy. This work proposes a novel neural architecture search called neural channel expansion that adjusts the network structure to alleviate accuracy degradation from ultra-low uniform-precision quantization. The proposed method selectively expands channels for the quantization sensitive layers while satisfying hardware constraints (e.g., FLOPs, PARAMs). Based on in-depth analysis and experiments, we demonstrate that the proposed method can adapt several popular networks channels to achieve superior 2-bit quantization accuracy on CIFAR10 and ImageNet. In particular, we achieve the best-to-date Top-1/Top-5 accuracy for 2-bit ResNet50 with smaller FLOPs and the parameter size.