MixTTA Enhances Model Adaptation to Data Shifts

A new research paper introduces MixTTA, a novel plug-in module designed to significantly improve Test-Time Adaptation (TTA) for deployed machine learning models. Current TTA methods often rely on adjusting affine parameters in normalization layers, which primarily handle axis-aligned scaling and shifting. This approach struggles with more complex, cross-channel structural changes that arise from distribution shifts in real-world data. MixTTA addresses this by integrating a low-rank cross-channel transformation into normalization layers, facilitating inter-channel mixing. The module also incorporates Decoupling Projection to ensure strict separation from the diagonal affine path and Spectral Projection to prevent rank-1 collapse, enhancing stability under non-stationary test streams. Experiments demonstrate that MixTTA consistently improves performance over strong baselines in various TTA settings and reduces adaptation failures in challenging conditions. The source code is publicly available.