Research map / Label-efficient and robust learning / Test-time adaptation

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

NeurIPS 2025 · Label-efficient and robust learning › Test-time adaptation · approach: Realistic and stable test-time adaptation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SNAP, a sparse test-time adaptation framework that pairs class/domain-representative memory sampling (CnDRM) with inference-only batch-aware memory normalization (IoBMN) to cut TTA latency on edge devices while keeping accuracy when adapting on only 1–50% of batches.

Problem

Existing test-time adaptation methods adapt at every batch with high computation and latency, which is impractical on resource-constrained edge devices; naively reducing adaptation frequency causes large accuracy drops.

Approach

Select update samples that are both high-confidence and close to the running domain centroid (Wasserstein distance on early-layer normalization statistics) to store in a batch-sized memory, and at inference re-normalize with soft-shrunk memory statistics toward the current batch statistics, so models can adapt sparsely without losing accuracy.

Key results

• On Raspberry Pi 4, SNAP reduces CoTTA latency by up to 93.12% at adaptation rate 0.1 on CIFAR10-C with no loss in accuracy (per paper claims).
• Integrated with five TTA methods, SNAP keeps accuracy drop below 3.3% across adaptation rates 1%–50%.
• On CIFAR100-C at AR=0.1, Tent+SNAP reaches 55.84% average accuracy versus 55.76% for full-adaptation Tent, and CoTTA+SNAP gets 50.52% vs 49.39% for full CoTTA.

Measured effects reported in the paper

Surprising findings

• Low-entropy sample filtering performed worse than random selection under sparse adaptation rates.
• At higher adaptation rates (0.3–0.5), SNAP can surpass fully adapted baselines.

Limitations

• Fixed adaptation rate and fixed confidence threshold; dynamic tuning not explored.
• Latency gains come from sparse batch-level updates rather than reducing backpropagation delay per step due to PyTorch constraints.

Experimental setup

Datasets: CIFAR10-C CIFAR100-C ImageNet-C ImageNet-R ImageNet-Sketch

Benchmarks: CIFAR10-C CIFAR100-C ImageNet-C ImageNet-R ImageNet-Sketch

Metrics: classification accuracy latency per batch peak memory

Baselines: Tent CoTTA EATA SAR RoTTA Naive STTA NOTE MECTA

Builds on: Tent EATA SAR CoTTA RoTTA Moderate Coreset NOTE

Closest papers by meaning

• A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
• Buffer layers for Test-Time Adaptation (NeurIPS 2025 · similarity 0.88)
• TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices (NeurIPS 2024 · similarity 0.88)
• Unraveling Batch Normalization for Realistic Test-Time Adaptation (AAAI 2024 · similarity 0.88)
• ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.88)
• Smoothing the Shift: Towards Stable Test-Time Adaptation under Complex Multimodal Noises (ICLR 2025 · similarity 0.88)