Research map / Optimization / Implicit bias and generalization
On the Dynamics Under the Unhinged Loss and Beyond
Xiong Zhou, Xianming Liu, Hanzhang Wang, Deming Zhai, Junjun Jiang, Xiangyang Ji
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 |
|---|---|---|---|
| add explicit feature regularization (lambda=5e-5) (ablation) | CE · Long-tailed CIFAR-10, ratio 100 | accuracy: 67.81 -> 69.74 | Improved |
| add explicit feature regularization (lambda=5e-5) (ablation) | CE · Long-tailed CIFAR-100, ratio 20 | accuracy: 42.96 -> 54.72 | Improved |
| add explicit feature regularization (lambda=1e-5) (ablation) | CE · Long-tailed CIFAR-100, ratio 100 | accuracy: 33.37 -> 36.61 | Improved |
| add explicit feature regularization (lambda=1e-6) (ablation) | unhinged loss · Long-tailed CIFAR-10, ratio 100 | accuracy: 72.60 -> 73.71 | Improved |
| add explicit feature regularization (lambda=1e-7) (ablation) | unhinged loss · Long-tailed CIFAR-10, ratio 100 | accuracy: 72.60 -> 71.85 | Hurt |
| add explicit feature regularization (lambda=5e-6) (ablation) | unhinged loss · Long-tailed CIFAR-10, ratio 100 | accuracy: 72.60 -> 71.13 | Hurt |
| add explicit feature regularization (lambda=5e-7) (ablation) | unhinged loss · Long-tailed CIFAR-100, ratio 100 | accuracy: 39.20 -> 41.85 | Improved |
| add explicit feature regularization (lambda=5e-6 or 1e-5) (ablation) | CE · balanced CIFAR-10, ratio 1 | accuracy: 95.28 -> 95.27 | No clear effect |
| add explicit feature regularization (lambda=1e-5) (ablation) | CE · balanced CIFAR-100, ratio 1 | accuracy: 75.42 -> 77.34 | Improved |
| add explicit feature regularization (lambda=5e-6) (ablation) | unhinged loss · balanced CIFAR-100, ratio 1 | accuracy: 75.37 -> 77.20 | Improved |
| add explicit feature regularization (EFL) (ablation) | CE · OOD detection CIFAR-10 vs SVHN, energy-based | FPR95: 43.04 -> 27.87 | Improved |
| add explicit feature regularization (EFL) (ablation) | unhinged loss · OOD detection CIFAR-10 vs SVHN, energy-based | FPR95: 53.33 -> 27.24 | Improved |
Closest papers by meaning
- Dichotomy of Feature Learning and Unlearning: Fast-Slow Analysis on Neural Networks with Stochastic Gradient Descent (ICML 2026 · similarity 0.89)
- Feature Learning Dynamics in Infinite-Depth 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)
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.87)
- Neural Collapse Beyond the Unconstrained Features Model: Landscape, Dynamics, and Generalization in the Mean-Field Regime (ICML 2025 · similarity 0.87)
- Global Convergence and Rich Feature Learning in $L$-Layer Infinite-Width Neural Networks under $\mu$ Parametrization (ICML 2025 · similarity 0.86)
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.
