Research map / Label-efficient and robust learning / Recognition and long-tail
Neural Collapse Inspired Feature Alignment for Out-of-Distribution Generalization
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 NCFAL, which alternates environment partitioning via per-environment local models and learning input masks whose semantic features collapse onto a fixed simplex ETF, improving OOD generalization on four datasets.
Problem
Spurious background correlations cause encoders to rely on non-semantic features, hurting out-of-distribution generalization, especially when environment labels are unavailable.
Approach
Uses neural collapse as an alignment target: a fixed simplex ETF classifier guides learned semantic masks so that invariant features from all environments collapse to the same class prototypes, while environments are inferred automatically by maximum-likelihood assignment with per-environment local models.
Key results
- Ours achieves 66.4% on ColoredMNIST, 58.0% on ColoredCOCO, 33.3% on COCOPlaces and 85.4% on NICO, outperforming ERM (51.5, 45.4, 20.1, 73.6) and all 16 listed OOD baselines.
- Without environment labels (w/o env), the method reaches 66.9%, 56.9%, 36.7% and 85.9% on the four datasets.
- Feature-level masking reaches 63.9% on ColoredCOCO and 43.7% on COCOPlaces, higher than pixel-level masking on those datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NCFAL instead of the strongest baseline, IRM | ColoredMNIST OOD · ColoredMNIST OOD | accuracy (%): 60.3 ± 2.8 -> 66.4 ± 0.2 | Helped |
| use NCFAL instead of the strongest baseline, CausIRL CORAL | ColoredCOCO OOD · ColoredCOCO OOD | accuracy (%): 51.5 ± 1.1 -> 58.0 ± 0.2 | Helped |
| use NCFAL instead of the strongest baseline, IRM | COCOPlaces OOD · COCOPlaces OOD | accuracy (%): 27.1 ± 0.9 -> 33.3 ± 1.7 | Helped |
| use NCFAL instead of the strongest baseline, MMD | NICO OOD · NICO OOD | accuracy (%): 78.2 ± 1.2 -> 85.4 ± 0.6 | Helped |
| automatically partition environments instead of using given environments (ablation) | NCFAL · ColoredMNIST OOD | accuracy (%): 66.4 ± 0.2 -> 66.9 ± 2.4 | No clear effect |
| automatically partition environments instead of using given environments (ablation) | NCFAL · ColoredCOCO OOD | accuracy (%): 58.0 ± 0.2 -> 56.9 ± 1.1 | Hurt |
| automatically partition environments instead of using given environments (ablation) | NCFAL · COCOPlaces OOD | accuracy (%): 33.3 ± 1.7 -> 36.5 ± 0.8 | Helped |
| apply the mask at the feature level instead of the pixel level (ablation) | NCFAL · ColoredMNIST OOD | accuracy (%): 66.4 ± 0.2 -> 58.7 ± 2.8 | Hurt |
| apply the mask at the feature level instead of the pixel level (ablation) | NCFAL · ColoredCOCO OOD | accuracy (%): 58.0 ± 0.2 -> 63.9 ± 0.5 | Helped |
| apply the mask at the feature level instead of the pixel level (ablation) | NCFAL · COCOPlaces OOD | accuracy (%): 33.3 ± 1.7 -> 43.7 ± 0.7 | Helped |
| replace maximum-likelihood environment assignment with random environment splits (ablation) | NCFAL · ColoredMNIST OOD | accuracy: 0.669 -> 0.417 | Hurt |
| replace maximum-likelihood environment assignment with random environment splits (ablation) | NCFAL · COCOPlaces OOD | accuracy: 0.367 -> 0.178 | Hurt |
Surprising findings
- Feature-level masks outperform pixel-level masks on ColoredCOCO and COCOPlaces, which the authors attribute to high-dimensional feature representations.
- Automatic environment partitioning without labels (w/o env) matches or exceeds the labeled-environment setting on several datasets.
Limitations
- Environment partitioning is heuristic and iterative; the authors note more effective and efficient partitioning techniques are needed.
- Evaluation is limited to four datasets, mostly small synthetic-spurious benchmarks with ResNet8/CNN backbones.
Experimental setup
Datasets: ColoredMNIST ColoredCOCO COCOPlaces NICO
Benchmarks: ColoredMNIST ColoredCOCO COCOPlaces NICO
Metrics: OOD accuracy (%) F-norm distance to simplex ETF
Baselines: ERM IRM VREx GroupDRO MLDG MMD IGA SANDMask Fish CDANN TRM IB ERM
Builds on: IRM neural collapse DomainBed
Closest papers by meaning
- Generalized Neural Collapse for a Large Number of Classes (ICML 2024 · similarity 0.86)
- Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learning (ICML 2025 · similarity 0.86)
- Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model (NeurIPS 2025 · similarity 0.86)
- Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition (AAAI 2025 · similarity 0.85)
- Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize (ICML 2024 · similarity 0.85)
- Pursuing Feature Separation based on Neural Collapse for Out-of-Distribution Detection (ICLR 2025 · similarity 0.84)
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.