用奖励机制引导扩散过程,提升分类模型在分布偏移下的鲁棒性。
SteeringTTA: Guiding Diffusion Trajectories for Robust Test-Time-Adaptation
- 基于费曼-卡茨理论,通过伪标签奖励引导多条扩散轨迹。
- 在ImageNet-C上无需模型更新即超越基线,对多种噪声类型通用。
- 仅需推理阶段操作,适合部署时快速适应新数据分布。
测试时自适应(TTA)旨在通过使用无标签测试数据来修正深度模型在分布偏移下的性能下降问题。现有的仅输入的基于扩散的TTA方法虽能提升对各类噪声的鲁棒性,但依赖梯度引导,限制了跨扰动类型的探索与泛化能力。本文提出SteeringTTA,一种仅需推理的框架,将费曼-卡茨引导机制引入扩散输入自适应中,以伪标签生成奖励信号。该方法维护多条粒子轨迹,由累积top-K概率与熵调度共同驱动,实现探索与置信度的平衡。在ImageNet-C数据集上,SteeringTTA在不进行任何模型更新或源数据访问的情况下,持续优于基线方法。
原文摘要 · Abstract (English)
Test-time adaptation (TTA) aims to correct performance degradation of deep models under distribution shifts by updating models or inputs using unlabeled test data. Input-only diffusion-based TTA methods improve robustness for classification to corruptions but rely on gradient guidance, limiting exploration and generalization across distortion types. We propose SteeringTTA, an inference-only framework that adapts Feynman-Kac steering to guide diffusion-based input adaptation for classification with rewards driven by pseudo-label. SteeringTTA maintains multiple particle trajectories, steered by a combination of cumulative top-K probabilities and an entropy schedule, to balance exploration and confidence. On ImageNet-C, SteeringTTA consistently outperforms the baseline without any model updates or source data.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。