Research map / Label-efficient and robust learning / Recognition and long-tail
Inducing Neural Collapse in Imbalanced Learning: Do We Really Need a Learnable Classifier at the End of Deep Neural Network?
Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, Dacheng Tao
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 |
|---|---|---|---|
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | ResNet · CIFAR-10, τ=0.005 | accuracy (%): 67.3 -> 71.9 | Improved |
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | ResNet · CIFAR-100, τ=0.005 | accuracy (%): 38.7 -> 40.9 | Improved |
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | ResNet · SVHN, τ=0.01 | accuracy (%): 40.9 -> 45.7 | Improved |
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | ResNet · STL-10, τ=0.01 | accuracy (%): 37.9 -> 37.2 | Hurt |
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | DenseNet · CIFAR-100, τ=0.005 | accuracy (%): 40.3 -> 40.1 | Hurt |
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | ResNet-50 · ImageNet-LT, 90 epochs | accuracy (%): 34.6 -> 41.8 | Improved |
| Replace the learnable classifier and CE with a fixed ETF classifier and DR loss | ResNet-34 · CUB-200-2011 | accuracy (%): 82.2 -> 83.0 | Improved |
| Replace the learnable classifier with a fixed ETF classifier while retaining CE (ablation) | ResNet · CIFAR-10, τ=0.005, without Mixup | accuracy (%): 66.1 -> 60.4 | Hurt |
| Replace CE with DR while using the fixed ETF classifier (ablation) | ResNet · CIFAR-10, τ=0.005, without Mixup | accuracy (%): 60.4 -> 68.4 | Improved |
| Add class-distribution weighting to CE (ablation) | Learnable classifier + CE · CIFAR-10, τ=0.005, without Mixup | accuracy (%): 66.1 -> 66.8 | Improved |
| Use Mixup with ETF classifier + DR (ablation) | ResNet · CIFAR-10, τ=0.005 | accuracy (%): 68.4 -> 71.9 | Improved |
| Use Mixup with ETF classifier + CE (ablation) | ResNet · CIFAR-10, τ=0.01 | accuracy (%): 72.9 -> 67.0 | Hurt |
Closest papers by meaning
- Neural Collapse for Cross-entropy Class-Imbalanced Learning with Unconstrained ReLU Features Model (ICML 2024 · similarity 0.94)
- Guiding Neural Collapse: Optimising Towards the Nearest Simplex Equiangular Tight Frame (NeurIPS 2024 · similarity 0.92)
- Neural Collapse To Multiple Centers For Imbalanced Data (NeurIPS 2024 · similarity 0.90)
- Understanding Representation of Deep Equilibrium Models from Neural Collapse Perspective (NeurIPS 2024 · similarity 0.88)
- Distribution Alignment Optimization through Neural Collapse for Long-tailed Classification (ICML 2024 · similarity 0.88)
- Wide Neural Networks Trained with Weight Decay Provably Exhibit Neural Collapse (ICLR 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.
