Research map / Label-efficient and robust learning / Domain adaptation and generalization
Limitations of Post-Hoc Feature Alignment for Robustness
Collin Burns, Jacob Steinhardt
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Re-estimate BatchNorm statistics on target data | AdaBN on CIFAR-10-C · CIFAR-10-C | accuracy: 72.3 -> 83.6 | Improved |
| Re-estimate BatchNorm statistics on target data | AdaBN on ImageNetV2 · ImageNetV2 | accuracy: 63.2 -> 60.9 | Hurt |
| Re-estimate BatchNorm statistics on target data | AdaBN on Stylized ImageNet · Stylized ImageNet | accuracy: 7.1 -> 10.2 | Improved |
| Re-estimate BatchNorm statistics on target data | AdaBN on black-border shifts · CIFAR-10 and TinyImageNet | accuracy: 65.0 -> 52.5; 22.6 -> 11.8 | Hurt |
| Use training augmentations when estimating target BatchNorm statistics | AdaBN on clean CIFAR-10 and TinyImageNet · CIFAR-10 and TinyImageNet | accuracy: +2.0; +3.7 | Improved |
| Use training augmentations when estimating target BatchNorm statistics | AdaBN on ImageNet-C and Stylized ImageNet · ImageNet-C and Stylized ImageNet | accuracy: -3.6; -1.8 | Hurt |
| Use training augmentations when estimating target BatchNorm statistics | AdaBN across reported datasets · Reported datasets | accuracy: 94.8 (+2.0); 64.0 (+3.7); 76.0 (+0.4); 86.7 (+3.1); 41.8 (+1.7); 43.3 (-3.6); 63.8 (+2.9); 8.4 (-1.8) | Mixed |
| Exclude the final BatchNorm layers from alignment (ablation) | AdaBN under CIFAR-10-C label shift · CIFAR-10-C | accuracy | Improved |
| Exclude the first BatchNorm layers from alignment (ablation) | AdaBN under CIFAR-10-C label shift · CIFAR-10-C | accuracy | No clear effect |
| Apply AdaBN to a smaller subset of target classes (ablation) | AdaBN · CIFAR-10-C | accuracy: below 50% | Hurt |
| Re-estimate BatchNorm statistics on target data | AdaBN · CIFAR-10-C, TinyImageNet-C, ImageNet-C, and Stylized ImageNet | ECE: 21.5 -> 11.3; 25.0 -> 15.2; 12.0 -> 5.2; 30.9 -> 12.9 | Improved |
| Re-estimate BatchNorm statistics on target data | AdaBN · ImageNetV2 | ECE: 10.6 -> 10.3 | Improved |
Closest papers by meaning
- Towards More Robust Interpretation via Local Gradient Alignment (AAAI 2023 · similarity 0.88)
- Distributionally Robust Post-hoc Classifiers under Prior Shifts (ICLR 2023 · similarity 0.87)
- Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks (ICLR 2025 · similarity 0.86)
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 2025 · similarity 0.85)
- Domain-Aware Fine-Tuning: Enhancing Neural Network Adaptability (AAAI 2024 · similarity 0.85)
- Sufficient Invariant Learning for Distribution Shift (CVPR 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.
