Research map / Label-efficient and robust learning / Test-time adaptation
Un-Mixing Test-Time Normalization Statistics: Combatting Label Temporal Correlation
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 UnMix-TNS, a test-time normalization layer that mixes each test instance's statistics with K online-updated BN statistics components, so BN statistics stay unbiased under temporally correlated (non-i.i.d.) test batches.
Problem
Test-time BN adaptation assumes i.i.d. test batches with respect to unknown labels; under temporally label-correlated streams the batch statistics become skewed and TTA methods fail.
Approach
Decompose stored BN statistics into K components, mix each instance's statistics with these components via cosine-similarity softmax assignment probabilities, and update components online with the most similar test instances, simulating i.i.d. normalization statistics without a memory bank.
Key results
- UnMix-TNS reduces average single-domain error on CIFAR10-C/CIFAR100-C/ImageNet-C to 45.6% vs 56.6-57.3% for other normalization baselines.
- Combining ROID with UnMix-TNS lowers mean single-domain error from 77.8% to 32.0% on the corruption benchmarks.
- UnMix-TNS with LAME reaches 33.7% mean single-domain error versus 48.4% for LAME alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UnMix-TNS instead of IABN | test-time normalization · CIFAR10-C single and continual | classification error (%): 29.1 -> 27.0; 29.1 -> 26.8 | Helped |
| Use UnMix-TNS instead of α-BN | test-time normalization · ImageNet-VID-C and LaSOT-C mean | classification error (%): 74.4 -> 68.6; 84.8 -> 81.1 | Helped |
| Add UnMix-TNS to ROID | ROID · CIFAR10-C, CIFAR100-C, and ImageNet-C single and continual | classification error (%): +45.8% or more stated; means: 77.8 -> 32.0 and 77.8 -> 31.8 | Helped |
| Add UnMix-TNS to TENT | TENT · CIFAR10-C single | classification error (%): 76.0 -> 27.0 | Helped |
| Add UnMix-TNS to CoTTA | CoTTA · CIFAR10-C single | classification error (%): 77.2 -> 49.1 | Helped |
| Add UnMix-TNS to LAME | LAME · mixed CIFAR100-C and ImageNet-C | classification error (%): 3.8 -> 4.2; 65.0 -> 68.5 | Hurt |
| Add UnMix-TNS to NOTE | NOTE · CIFAR10-C single and mixed | classification error (%): 26.0 -> 26.7; 36.1 -> 52.6 | Mixed |
| Add UnMix-TNS to RoTTA | RoTTA · CIFAR10-C single | classification error (%): 27.7 -> 26.8 | Helped |
| Add UnMix-TNS to LAME | LAME · ImageNet-VID-C and LaSOT-C mean | classification error (%): 68.0 -> 64.8; 78.9 -> 78.1 | Helped |
| Increase correlation by decreasing Dirichlet concentration δ (ablation) | TBN, α-BN, and RBN · CIFAR100-C | classification error (%) | Hurt |
| Vary batch size from 1 to 256 (ablation) | UnMix-TNS · temporally correlated continual and mixed domain streams | classification error (%) | No clear effect |
| Increase K above 16 (ablation) | UnMix-TNS · CIFAR10-C single and continual domain | classification error (%) | Hurt |
Surprising findings
- Under mixed-domain adaptation, the IABN and alpha-BN-style methods can be competitive, and LAME alone is best on some mixed-domain CIFAR100-C and ImageNet-C settings where UnMix-TNS does not win.
- Source-statistics-based methods degrade when source-domain test data reappears, and larger batch sizes do not fix this for TBN.
Limitations
- Not designed for test batches containing wide ranges of diverse or outlier instances.
- Number of components K must be tuned and varies by dataset and scenario.
- Only evaluated on classification; BN-based segmentation models left for future work.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C ImageNet-C DomainNet-126 ImageNet-VID LaSOT
Benchmarks: CIFAR10-C CIFAR100-C ImageNet-C DomainNet-126 ImageNet-VID-C LaSOT-C
Metrics: online classification error rate
Baselines: TBN alpha-BN IABN RBN TENT CoTTA LAME NOTE RoTTA ROID
Builds on: test-time BN alpha-BN NOTE RoTTA Momentum BN
Closest papers by meaning
- Unraveling Batch Normalization for Realistic Test-Time Adaptation (AAAI 2024 · similarity 0.89)
- Buffer layers for Test-Time Adaptation (NeurIPS 2025 · similarity 0.87)
- Test-Time Adaptation by Causal Trimming (NeurIPS 2025 · similarity 0.85)
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.85)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.85)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.85)
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.