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

Position

First author project in KAIST Mobile Intelligence & Interaction Lab

Project Goals & Works

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

Key Results

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

Links

• Website: https://miil.kaist.ac.kr/projects/snap
• arXiv: https://arxiv.org/abs/2511.15276
• Code: https://github.com/chahh9808/SNAP