Reproduce “DELTA: degradation-free fully test-time adaptation” 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
Bowen Zhao, Chen Chen, Shu-Tao Xia
DELTA, an ICLR 2023 paper from researchers at Tsinghua University, Tencent, and OPPO, introduces a plug-in framework to enhance fully test-time adaptation by addressing inaccurate normalization statistics and biased parameter optimization. It consistently improves model performance across diverse real-world scenarios, including time-dependent and class-imbalanced data streams, on datasets like ImageNet-C and a new YTBB-sub video dataset.
Key insights
- Mini-batch statistics used in test-time batch normalization are often noisy and unrepresentative of true population statistics, leading to significant performance degradation in time-dependent data streams.
- Self-learning based test-time adaptation methods tend to amplify existing prediction biases, causing parameters to optimize towards dominant classes and negatively impacting performance in class-imbalanced scenarios.
- Achieving truly robust fully test-time adaptation requires simultaneously addressing both the stability of normalization statistics and the balance of parameter optimization across diverse and dynamic data streams.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.