SNAP-TTA: Sparse Test-Time Adaptation for Latency-Sensitive Applications
Keywords: Test-Time Adaptation, Unsupervised Domain Adaptation
Abstract: Test-Time Adaptation (TTA) methods use unlabeled test data to dynamically adjust models in response to distribution changes. However, existing TTA methods are not tailored for practical use on edge devices with limited computational capacity, resulting in a latency-accuracy trade-off. To address this problem, we propose SNAP-TTA, a sparse TTA framework that significantly reduces adaptation frequency and data usage, delivering latency reductions proportional to adaptation rate. It achieves competitive accuracy even with an adaptation rate as low as 0.01, demonstrating its ability to adapt infrequently while utilizing only a small portion of the data relative to full adaptation. Our approach involves (i) Class and Domain Representative Memory (CnDRM), which identifies key samples that are both class-representative and domain-representative to facilitate adaptation with minimal data, and (ii) Inference-only Batch-aware Memory Normalization (IoBMN), which leverages representative samples to adjust normalization layers on-the-fly during inference, aligning the model effectively to changing domains. When combined with five state-of-the-art TTA algorithms, SNAP-TTA maintains the performances of these methods even with much-reduced adaptation rates from 0.01 to 0.5, making it suitable for edge devices serving latency-sensitive applications.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 14267
Loading