WISE-ATTA: When to Ask for Labels in Budgeted Active Test-Time Adaptation
WISE-ATTA schedules scarce labels during test-time adaptation and matches prior methods with fewer annotations.
The paper defines budgeted active test-time adaptation, where labels can be requested for only a fraction of incoming test batches. WISE-ATTA allocates that limited supervision over time using lightweight signals computed online, then selects samples whose adaptation has not yet converged so one labeled example can drive an update. On synthetic ImageNet-C corruptions and natural shifts in ImageNet-R, ImageNet-K, and ImageNet-A, it matches or improves on recent methods while using substantially fewer labels.
- Labels are available for only a fraction of test batches.
- Lightweight online signals decide when supervision is worth spending.
- Drift-based selection can update a model from one labeled sample.
- Results cover ImageNet-C plus ImageNet-R, ImageNet-K, and ImageNet-A.
From the original article
1 min read at huggingface.co
Active test-time adaptation (ATTA) improves robustness under distribution shift by updating a deployed model during inference while selectively querying supervision. However, most existing ATTA methods implicitly assume that supervision can be requested for every incoming test batch, which can incur substantial annotation cost over long test streams.
By Muhammad Huzaifa, Lea Schönherr, Thorsten Eisenhofer · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.