Research map / Label-efficient and robust learning / Recognition and long-tail
Maximum Class Separation as Inductive Bias in One Matrix
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| add fixed maximally separated matrix before softmax | ResNet-32 · CIFAR-10 imbalance 0.01 | accuracy: 56.98 -> 69.70 | Helped |
| add fixed maximally separated matrix before softmax | ResNet-32 · CIFAR-100 imbalance 0.01 | accuracy: 35.02 -> 38.85 | Helped |
| add fixed maximally separated matrix before softmax | ConvNet · CIFAR-100 balanced | accuracy: 56.70 -> 57.05 | No clear effect |
| make the maximum-separation matrix learnable (initialized with it) instead of fixed (ablation) | max separation matrix · CIFAR-10/100 long-tailed | accuracy | Hurt |
| add extra learnable FC layer with standard init instead of fixed matrix (ablation) | ResNet-32 · CIFAR-10/100 long-tailed | accuracy | Hurt |
| add fixed maximum separation on top of LDAM-SGD | LDAM-SGD · CIFAR-100 imbalance 0.01 | accuracy: 39.87 -> 42.02 | Helped |
| add fixed maximum separation on top of MiSLAS stage 2 | MiSLAS · CIFAR-100 imbalance 0.1 | accuracy: 61.93 -> 63.52 | Helped |
| add fixed maximum separation | ResNet-50 · ImageNet-LT | top-1 accuracy: 43.8 -> 47.3 | Helped |
| add fixed maximum separation | ResNet-152 · ImageNet | top-1 accuracy: 77.9 -> 78.5 | Helped |
| add fixed maximum separation | WideResNet energy score · CIFAR-100 vs SVHN | AUROC: 73.94 -> 85.42 | Helped |
| add fixed maximum separation | WideResNet energy score · CIFAR-100 vs Places 365 | FPR95: 80.87 -> 81.36 | Mixed |
| add fixed maximum separation | MLS open-set recognition · CIFAR10 | AUROC: 93.82 -> 95.30 | Helped |
Closest papers by meaning
- Separation and Concentration in Deep Networks (ICLR 2021 · similarity 0.88)
- Implicit Bias of Spectral Descent and Muon on Multiclass Separable Data (NeurIPS 2025 · similarity 0.85)
- Separation Power of Equivariant Neural Networks (ICLR 2025 · similarity 0.84)
- Training the Untrainable: Introducing Inductive Bias via Representational Alignment (NeurIPS 2025 · similarity 0.84)
- Tripod: Three Complementary Inductive Biases for Disentangled Representation Learning (ICML 2024 · similarity 0.84)
- Inductive biases of multi-task learning and finetuning: multiple regimes of feature reuse (NeurIPS 2024 · 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.