Reproduce “Un-Mixing Test-Time Normalization Statistics: Combatting Label Temporal Correlation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, Behzad Bozorgtabar
A system called UnMix-TNS, developed by researchers at EPFL and CHUV, addresses the degradation of deep learning models in real-world settings by maintaining multiple distinct normalization statistics components to counteract label temporal correlation in test-time adaptation. The method achieves average accuracy improvements of over 11% on corruption benchmarks and up to 60% when integrated with existing test-time adaptation techniques, while introducing only 0.15ms additional inference latency per image.
Key insights
- Label temporal correlation in real-world test data streams is a critical and often overlooked cause of performance degradation in Batch Normalization-based test-time adaptation.
- Decomposing and dynamically managing multiple distinct statistical profiles can effectively simulate i.i.d. conditions for normalization, even when actual batches are highly correlated.
- This 'unmixing' approach offers a generic, low-overhead way to enhance the robustness and stability of a wide range of existing test-time adaptation methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.