SNAP: Low-Latency Test-Time Adaptation with Sparse Updates

Wed, 01 Oct 2025 00:00:00 +0000

Position

First author project in KAIST Mobile Intelligence & Interaction Lab

Designed SNAP, a sparse Test-Time Adaptation (TTA) framework for latency-sensitive edge applications.
Proposed Class and Domain Representative Memory (CnDRM) to select a compact and informative subset of target samples for adaptation.
Proposed Inference-only Batch-aware Memory Normalization (IoBMN) to align normalization statistics at inference time with minimal overhead.
Integrated SNAP with five state-of-the-art TTA methods and validated consistent speedups with limited accuracy loss.

Reduced adaptation latency by up to 93.12% compared with baseline TTA pipelines.
Maintained competitive performance with less than 3.3% accuracy drop across adaptation rates from 1% to 50%.
Demonstrated strong practicality for resource-constrained, real-time on-device inference scenarios.