Reproduce “Evaluating Prediction-Time Batch Normalization for Robustness under Covariate Shift” 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
Zachary Nado, Shreyas Padhy, D. Sculley, Alexander D'Amour, Balaji Lakshminarayanan, Jasper Snoek
The authors evaluate prediction-time batch normalization (BN) for image recognition under covariate shift, recomputing activation statistics from unlabeled prediction batches without changing model weights. Across corruption benchmarks, it often improves calibration and robustness measures; on ImageNet-C, a vanilla ResNet-50 achieves 60.28% mCE. Results are not universal: some natural-shift and ensemble metrics worsen.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.