深度神经网络的低比特模型量化：综述
机器学习
2025-05-12 v1 人工智能
摘要
随着前所未有的快速发展，深度神经网络（DNNs）已深刻影响几乎所有领域。然而，其高昂的计算成本和模型规模在真实部署中通常难以接受。模型量化，一种有效的权重压缩技术，已成为整个部署流程中不可或缺的步骤。量化加速的本质是将连续浮点数转换为离散整数，这显著加速了内存I/O和计算，即加法和乘法。然而，由于精度损失，转换也会带来性能下降。因此，研究如何执行转换以及如何补偿信息损失已变得越来越流行和关键。本文综述了最近五年在DNN低比特量化方面的进展。我们讨论并比较了最先进的量化方法，并根据其核心技术将其分为8个主类和24个子类。此外，我们揭示了模型量化领域的潜在研究机会。精选的模型量化列表可在 https://github.com/Kai-Liu001/Awesome-Model-Quantization 获取。
引用
@article{arxiv.2505.05530,
title = {Low-bit Model Quantization for Deep Neural Networks: A Survey},
author = {Kai Liu and Qian Zheng and Kaiwen Tao and Zhiteng Li and Haotong Qin and Wenbo Li and Yong Guo and Xianglong Liu and Linghe Kong and Guihai Chen and Yulun Zhang and Xiaokang Yang},
journal= {arXiv preprint arXiv:2505.05530},
year = {2025}
}
备注
We have systematically collected and reviewed the state-of-the-art quantization methods from the past five years, categorizing them into eight distinct groups. A curated list of model quantization is provided at https://github.com/Kai-Liu001/Awesome-Model-Quantization