Research map / Label-efficient and robust learning / Recognition and long-tail
Neural Collapse by Design: Learning Class Prototypes on the Hypersphere
Panagiotis Koromilas, Theodoros Giannakopoulos, Mihalis Nicolaou, Yannis Panagakis
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 weights to the hypersphere (NormFace) instead of CE (ablation) | classifier learning · CIFAR-10/100, ImageNet-100/1K | top-1 accuracy: 75.4 -> 76.4 (ImageNet-1K) | Improved |
| use all batch samples as negatives (NTCE) instead of K class prototypes (ablation) | NormFace · ImageNet-1K | top-1 accuracy: 76.4 -> 76.7 | Improved |
| exclude same-class instances from the denominator (NONL) (ablation) | NTCE · CIFAR-100 | top-1 accuracy: 72.9 -> 73.6 | Improved |
| NONL instead of NTCE on ImageNet-1K (ablation) | NTCE/NONL · ImageNet-1K | top-1 accuracy: 76.7 vs 76.5 | No clear effect |
| replace linear probing with fixed class-mean prototypes | SCL · CIFAR-10, CIFAR-100, ImageNet-1K | top-1 accuracy: 75.1 -> 75.1 (ImageNet-1K) | No clear effect |
| replace linear probing with fixed class-mean prototypes | SCL · ImageNet-100 | top-1 accuracy: 84.8 -> 86.8 | Improved |
| replace linear probing with normalized linear probing (ablation) | SCL · CIFAR-10, CIFAR-100, ImageNet | top-1 accuracy: 73.9 -> 73.6 (CIFAR-100) | No clear effect |
| use NONL/NTCE instead of CE | classifier learning · CIFAR-10/100 training | NC effective rank and alignment: Inter ER 8.6/57.1 -> 9.0/99.0 | Improved |
| train SCL with linear probing vs fixed prototypes (ablation) | SCL · CIFAR-10/100 | NC3 weight-feature alignment: 0.99/1.03 -> 0.00/0.00 | Improved |
| use NONL instead of CE for pretraining | ResNet50 · Transfer learning (8 datasets) | mean accuracy: 67.0 -> 70.7 | Improved |
| use ETF + DR classifier (ablation) | ResNet50 · Transfer learning (8 datasets) | mean accuracy: 67.0 -> 54.6 | Hurt |
| use NONL instead of CE | ResNet18 · CIFAR-100-LT (ratio 0.01) | accuracy: 37.4 -> 40.0 | Improved |
Closest papers by meaning
- Engineering the Neural Collapse Geometry of Supervised-Contrastive Loss (Student Abstract) (AAAI 2024 · similarity 0.91)
- Neural Collapse Beyond the Unconstrained Features Model: Landscape, Dynamics, and Generalization in the Mean-Field Regime (ICML 2025 · similarity 0.89)
- Symmetric Neural-Collapse Representations with Supervised Contrastive Loss: The Impact of ReLU and Batching (ICLR 2024 · similarity 0.88)
- Guiding Neural Collapse: Optimising Towards the Nearest Simplex Equiangular Tight Frame (NeurIPS 2024 · similarity 0.88)
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.87)
- BCE vs. CE in Deep Feature Learning (ICML 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.
