Research map / Label-efficient and robust learning / Domain adaptation and generalization
Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration
Cian Eastwood, Ian Mason, Chris Williams, Bernhard Schölkopf
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| Use bottom-up training rather than adapting all layers simultaneously (ablation) | FR · EMNIST-DA | accuracy: 74.4±0.8 -> 86.1±0.1 | Improved |
| Use bottom-up training rather than adapting all layers simultaneously (ablation) | FR · CIFAR-10-C | accuracy: 87.2 ± 0.7 -> 89.4 ± 0.2 | Improved |
| Use bottom-up training rather than adapting all layers simultaneously (ablation) | FR · CIFAR-100-C | accuracy: 65.5 ± 0.2 -> 68.5 ± 0.2 | Improved |
| Use bottom-up training rather than adapting all layers simultaneously (ablation) | FR · Camelyon17 | accuracy: 84.6±0.6 -> 84.5±0.8; 86.0±1.1 -> 87.0±1.2; 89.5±0.4 -> 89.7±0.5 | Mixed |
| Add marginal logit-distribution alignment (ablation) | Feature Restoration · CIFAR-10-C | accuracy | No clear effect |
| Add marginal logit-distribution alignment (ablation) | Feature Restoration · CIFAR-100-C | accuracy | Improved |
| Use symmetric KL divergence instead of asymmetric KL divergence (ablation) | Feature Restoration · CIFAR-10-C | accuracy | No clear effect |
| Use symmetric KL divergence instead of asymmetric KL divergence (ablation) | Feature Restoration · CIFAR-100-C | accuracy | Improved |
| Align source and target features rather than only replacing batch-normalization statistics | Feature Restoration · EMNIST-DA | accuracy: 46.2±1.1 -> 74.4±0.8 | Improved |
| Use feature-distribution restoration instead of entropy-minimizing SHOT | BUFR · EMNIST-DA | ECE: 29.6±3.7 -> 4.7±0.2 | Improved |
| Adapt features with BUFR instead of entropy-minimizing SHOT | BUFR · EMNIST-DA severe | accuracy: 55.1±23.5 -> 84.6±0.2 | Improved |
| Correct label shift before applying feature restoration (ablation) | FR · VisDA-C with ImageNet pretraining | average accuracy: 73.7 -> 82.8 | Improved |
Closest papers by meaning
- Robust Nearest Neighbors for Source-Free Domain Adaptation under Class Distribution Shift (ECCV 2024 · similarity 0.88)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.88)
- Unraveling the Mysteries of Label Noise in Source-Free Domain Adaptation: Theory and Practice (TPAMI 2025 · similarity 0.88)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · similarity 0.87)
- Discriminative Pattern Calibration Mechanism for Source-Free Domain Adaptation (CVPR 2024 · similarity 0.86)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.86)
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.
