Reproduce “AETTA: Label-Free Accuracy Estimation for 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
Taeckyung Lee, Sorn Chottananurak, Taesik Gong, Sung-Ju Lee
AETTA estimates unlabeled batch accuracy during test-time adaptation by comparing model predictions with stochastic dropout predictions. It weights disagreement using batch entropy to account for skewed class predictions. Under calibration and dropout-independence assumptions, expected disagreement relates to test error. On continual CIFAR100-C, MAE was 5.74 ± 0.13% versus 15.18 ± 0.09% for AdvPerturb, averaged across six methods, 15 corruptions, and three seeds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.