Research map / Label-efficient and robust learning / Recognition and long-tail
Neural Collapse with Normalized Features: A Geometric Analysis over the Riemannian Manifold
Can Yaras, Peng Wang, Zhihui Zhu, Laura Balzano, Qing Qu
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 |
|---|---|---|---|
| Normalize features and classifiers to the sphere rather than using no normalization | UFM · UFM, 100 classes, 5 samples per class | Average accuracy: 49.9% ± 2.39% -> 100.0% ± 0.00% | Improved |
| Normalize features and classifiers to the sphere rather than using no normalization | UFM · UFM, 100 classes, 5 samples per class | Average CE loss: 1.63 ± 0.03 -> 3.84 ± 0.00 | Hurt |
| Use feature normalization instead of regularization | ResNet-18 · CIFAR100 test split | Test accuracy: 55.3% -> 58.6% | Improved |
| Use feature normalization instead of regularization | ResNet-18 · CIFAR100 test split | Test NC1: 3.838 -> 3.143 | Improved |
| Use feature normalization instead of regularization | ResNet-50 · CIFAR100 test split | Test accuracy: 48.9% -> 56.4% | Improved |
| Use feature normalization instead of regularization | ResNet-50 · CIFAR100 test split | Test NC1: 4.486 -> 3.127 | Improved |
| Increase network width | ResNet-18 with normalized features and classifiers · CIFAR10 with random labels | NC metrics | Improved |
| Use feature normalization instead of regularization | ResNet-18 · Reduced CIFAR100 training set | Training speed to perfect classification: 10-20 epochs sooner | Improved |
| Use feature normalization instead of regularization | ResNet-50 · Reduced CIFAR100 training set | Training accuracy: above 90% | Improved |
| Set temperature in the range around 1 to 10, with values close to 5 (ablation) | Feature-normalized ResNet-18 · CIFAR-10 | NC metrics | Improved |
| Use a larger temperature outside the range that achieves perfect classification (ablation) | Feature-normalized ResNet-18 · CIFAR-10 | Training accuracy: from about 0.1 to 10 | Hurt |
| Use label smoothing with sufficiently large temperature rather than ordinary CE (ablation) | Normalized UFM · Classification under label smoothing | Neural collapse | Hurt |
Closest papers by meaning
- Guiding Neural Collapse: Optimising Towards the Nearest Simplex Equiangular Tight Frame (NeurIPS 2024 · similarity 0.89)
- Generalized Neural Collapse for a Large Number of Classes (ICML 2024 · similarity 0.89)
- A Lie Group Approach to Riemannian Batch Normalization (ICLR 2024 · similarity 0.88)
- Understanding Matrix Function Normalizations in Covariance Pooling through the Lens of Riemannian Geometry (ICLR 2025 · similarity 0.88)
- Building Transformation Layers for Riemannian Neural Networks (NeurIPS 2026 · similarity 0.87)
- Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model (NeurIPS 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.
