Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Adaptation on Noisy Data via Model-Pruning-Based Filtering and Flatness-Aware Entropy Minimization
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 MoTTA, a buffer-based test-time adaptation method for noisy data streams that filters samples via output difference under structured pruning and updates the model with a flatness-aware entropy loss combining zeroth- and first-order flatness plus parameter regularization.
Problem
Test-time adaptation on unlabeled data streams that contain out-of-source-category noisy samples, where inaccurate predictions under domain shift degrade noisy-sample filtering and flatness-aware optimization only considers worst-case (zeroth-order) flatness.
Approach
Filter samples by the angular output difference before and after structured (L1-norm) channel pruning instead of raw prediction confidence, and optimize entropy with combined zeroth-order and first-order (gradient-norm) flatness approximated via Taylor expansion, plus a parameter regularizer toward the source weights.
Key results
- On ImageNet-C with SH noise, MoTTA reaches 53.4% average error vs. 56.6% for SoTTA, a 3.2-point reduction.
- On ImageNet-R with SH noise, MoTTA reaches 56.1% error vs. 57.6% for SoTTA; on ImageNet-K with SH, 66.0% vs. 67.5%.
- Ablation on ImageNet-R (SH) shows removing ODP-based filtering raises error from 56.1% to 57.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoTTA instead of the strongest reported baseline | TTA on ImageNet-C with noisy samples · ImageNet-C, SH and NINCO | classification error rate (%): 56.6 -> 53.4; 57.3 -> 54.2 | Helped |
| Use MoTTA instead of SoTTA | TTA on ImageNet-R and ImageNet-K with noisy samples · ImageNet-R and ImageNet-K, SH and NINCO | classification error rate (%): 57.6 -> 56.1; 58.4 -> 56.7; 67.5 -> 66.0; 67.5 -> 66.8 | Helped |
| Add ODP-based filtering (ablation) | MoTTA · ImageNet-R and ImageNet-K, SH and NINCO | classification error rate (%): 57.4 -> 56.1; 57.5 -> 56.7; 67.3 -> 66.0; 67.4 -> 66.8 | Helped |
| Add first-order flatness R1 (ablation) | MoTTA · ImageNet-R and ImageNet-K, SH and NINCO | classification error rate (%): 56.5 -> 56.1; 56.8 -> 56.7; 66.5 -> 66.0; 67.3 -> 66.8 | Helped |
| Add parameter regularization Rr (ablation) | MoTTA · ImageNet-R and ImageNet-K, SH and NINCO | classification error rate (%): 56.8 -> 56.1; 58.2 -> 56.7; 67.1 -> 66.0; 68.3 -> 66.8 | Helped |
| Add confidence-based filtering (ablation) | MoTTA · ImageNet-R and ImageNet-K, SH and NINCO | classification error rate (%): 57.2 -> 56.1; 57.9 -> 56.7; 68.1 -> 66.0; 68.1 -> 66.8 | Helped |
Surprising findings
- Unstructured pruning does not yield a clear bimodal ODP score distribution, so structured pruning is required for the filter to work.
- Pseudo-label baseline performs worst among methods, with errors above 85% on ImageNet-R and ImageNet-K.
Limitations
- Gains over SoTTA are modest on some datasets (e.g., SoTTA is second-best on several ImageNet-C corruption types and MoTTA does not win on all settings).
- Relies on a hand-set pruning ratio and thresholds (tau_ODP, tau_c) whose sensitivity is not extensively analyzed in the main text.
- Evaluation limited to image classification with two noisy datasets.
Experimental setup
Datasets: ImageNet-C ImageNet-R ImageNet-Sketch SSB-hard NINCO
Benchmarks: ImageNet-C ImageNet-R ImageNet-Sketch
Metrics: classification error rate
Baselines: Source PL BN Stats TENT CoTTA RoTTA SAR OSTTA UniEnt SoTTA
Builds on: TENT SoTTA SAM Gradient norm aware minimization EATA
Closest papers by meaning
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.90)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.88)
- Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation (ICML 2025 · similarity 0.88)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.88)
- SNAP: Low-Latency Test-Time Adaptation with Sparse Updates (NeurIPS 2025 · similarity 0.87)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.