arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2609.22238cs.LGcs.CV

任务感知的混合QUBO优化用于结构化神经网络剪枝

Task-Aware QUBO Allocation for Mixed-Precision Quantization

Osama Orabi, Artur Zagitov, Hadi Salloum, Viktor A. Lobachev, Yaroslav Kholodov

首次发表
浏览论文内容

中文总结 AI 辅助

提出混合QUBO框架,结合任务敏感性与滤波器交互进行结构化剪枝,通过二分搜索控制基数,并采用QUBO-张量列细化优化掩码,在SIDD去噪任务上超越基线。

中文摘要 AI 辅助

神经网络剪枝可以表述为一个组合优化问题,然而许多现有方法依赖于独立的滤波器重要性评分或简化的目标函数。在本工作中,我们提出了一种用于结构化滤波器剪枝的混合二次无约束二元优化(QUBO)框架,该框架将任务感知的敏感性信息与候选滤波器之间的交互相结合。该公式将一阶泰勒敏感性和权重-费舍尔敏感性纳入目标的线性部分,并可额外将激活相似性纳入二次交互项。为了在不引入显式二次基数惩罚的情况下控制目标剪枝基数,我们对容量激励进行二分搜索,以确定一个经验上能达到目标剪枝基数的系数。我们进一步研究了一种两阶段的QUBO-张量列细化策略,其中QUBO解初始化基于梯度的概率黑盒优化,以使用下游指标搜索改进的剪枝掩码。在SIDD图像去噪任务和Half-UNet模型上的实验表明,在所研究的剪枝目标下,混合QUBO相比评估的基于泰勒和基于L1的QUBO基线获得了更高的PSNR和SSIM。在固定数据集协议下进行多种子实验以评估鲁棒性,而受控子问题实验表明,随着组合问题规模的增大,张量列细化变得越来越有价值。结果支持混合QUBO作为所评估设置下的任务感知结构化剪枝框架,同时也强调了基于掩码剪枝的计算和部署局限性。

英文摘要

Mixed-precision quantization requires discrete allocation of weight and activation bit-widths, followed by recovery of the selected network. We develop a task-aware quadratic unconstrained binary optimization (QUBO) surrogate with separate weight and activation profiles, a bit-operation (BOP) cost, and selected structural priors. QUBO provides a network-wide allocation that can be refined through direct validation-based PROTES search. On a compact NAFBlock-based denoiser, the refined route achieves 37.192 dB after LSQ+ at 4.035\% routed-layer BOPs, versus 37.092 dB at 4.101\% for a HAWQ-style baseline. The repeated-search primary experiment shows that LSQ+ largely closes the quality gap between QUBO allocation and expensive direct refinement. An additional restoration architecture retains a larger recovered gain, indicating that refinement's value depends on architecture and recovery. We evaluate quality, achieved cost, routing stability and optimization expense together. Deployment measurements characterize a fake-quantized floating-point implementation; BOP reductions describe analytical allocation savings.

发表机构

  • Innopolis University(因诺波利斯大学)
  • Q Deep
  • Moscow Independent Research Institute of Artificial Intelligence (MIRAI)(莫斯科独立人工智能研究院(MIRAI))
  • Sirius University of Science and Technology(天狼星科技大学)

机构由 AI 辅助整理,请以论文原文为准。

补充信息

↑