Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization
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 MABN, which adapts only the BN affine parameters at test time with a label-independent self-supervised auxiliary branch, meta-trained via bi-level optimization so the adaptation aligns with the main task.
Problem
Test-time domain adaptation: adapting a source-trained model to unseen target domains using only a few unlabeled images, where prior methods update the whole network and entangle label and domain knowledge, and BN statistics estimated from few samples are unstable.
Approach
Freeze the weight matrix and source BN running statistics and adapt only the BN affine parameters (gamma, beta) using a BYOL-style SSL auxiliary branch; meta-train these parameters with MAML-style bi-level optimization over source domains as tasks so that the SSL adaptation improves the main classification/regression objective. The auxiliary branch is discarded at inference.
Key results
- On WILDS-iWildCam, MABN reaches 78.4% accuracy and 38.3 Macro F1, versus 77.2% / 34.0 for Meta-DMoE, and outperforms all compared methods on the eight WILDS metrics reported.
- On DomainNet, MABN averages 45.5% accuracy versus 44.2% for Meta-DMoE and 35.5% for ARM.
- Ablation on iWildCam shows adapting the whole BN layer degrades accuracy (70.5 to 68.2 without meta-training), while adapting only affine parameters with meta-auxiliary training reaches 78.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the SSL branch and joint training (ablation) | MABN training on iWildCam · iWildCam | accuracy: 68.7 -> 70.5 | Helped |
| Add the SSL branch and joint training (ablation) | MABN training on iWildCam · iWildCam | Macro F1: 31.3 -> 33.2 | Helped |
| Adapt by updating the whole BN layer instead of no adaptation (ablation) | MABN on iWildCam · iWildCam | accuracy: 70.5 -> 68.2 | Hurt |
| Adapt by updating only BN affine parameters instead of the whole BN layer (ablation) | MABN on iWildCam · iWildCam | accuracy: 68.2 -> 71.1 | Helped |
| Use meta-training rather than joint training with affine adaptation (ablation) | MABN without target adaptation on iWildCam · iWildCam | accuracy: 71.1 -> 74.7 | Helped |
| Add target adaptation to meta-trained affine parameters (ablation) | MABN on iWildCam · iWildCam | accuracy: 74.7 -> 78.4 | Helped |
| Use meta-training and adaptation while updating all parameters rather than affine parameters (ablation) | MABN on iWildCam · iWildCam | Macro F1: 38.3 -> 29.4 | Hurt |
| Use domain-matched adapted affine parameters rather than non-matched parameters (ablation) | MABN on target domains · iWildCam | accuracy: 72.39 -> 78.40 | Helped |
| Use domain-matched adapted affine parameters rather than non-matched parameters (ablation) | MABN on target domains · iWildCam | Macro F1: 33.32 -> 38.27 | Helped |
| Update only affine parameters instead of the entire BN layer (ablation) | TENT adaptation · iWildCam | accuracy: 33.27 -> 75.92 | Helped |
| Add MABN before TENT | MABN on iWildCam · iWildCam | accuracy: +1.28% | Helped |
| Add MABN before TENT | MABN on iWildCam · iWildCam | Macro F1: +0.58% | Helped |
Surprising findings
- Adapting the entire BN layer (including normalization statistics) on a few target images degrades performance relative to the joint-trained model, showing unstable few-shot statistics interfere with adaptation.
- Without meta-auxiliary training, adapting the jointly trained model with SSL sometimes drops performance below the no-adaptation baseline.
Limitations
- Adaptation relies on a meta-training phase over labeled source domains, so it requires source-domain labels and a bi-level optimization that adds training cost.
- Gains over strong baselines on some metrics are small (e.g., PovertyMap Pearson r, where ARM-BN is higher on worst-case metrics in the table), and results rely on 5 seeds with notable variance on some datasets.
Experimental setup
Datasets: iWildCam Camelyon17 RxRx1 FMoW PovertyMap DomainNet ImageNet-1K
Benchmarks: iWildCam Camelyon17 RxRx1 FMoW PovertyMap DomainNet
Metrics: Accuracy Macro F1 worst-case accuracy Pearson r worst-case Pearson r
Baselines: ERM CORAL Group DRO IRM ARM-CML ARM-BN ARM-LL Meta-DMoE PAIR ARM TENT
Builds on: ARM Meta-DMoE MAML BYOL TENT
Closest papers by meaning
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 2025 · similarity 0.89)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.88)
- Adapting to Distribution Shift by Visual Domain Prompt Generation (ICLR 2024 · similarity 0.88)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.88)
- Unraveling Batch Normalization for Realistic Test-Time Adaptation (AAAI 2024 · similarity 0.88)
- Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation (ICLR 2025 · 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.