SteeringTTA: Guiding Diffusion Trajectories for Robust Test-Time-Adaptation
This addresses robust test-time adaptation for classification tasks, particularly in handling distribution shifts like corruptions, but appears incremental as it builds upon existing diffusion-based methods.
The paper tackled the problem of performance degradation in deep models under distribution shifts by proposing SteeringTTA, an inference-only framework that guides diffusion-based input adaptation for classification without model updates or source data, achieving consistent outperformance over baselines on ImageNet-C.
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.