Research map / Label-efficient and robust learning / Recognition and long-tail
Cut your Losses with Squentropy
Like Hui, Mikhail Belkin, Stephen Wright
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 cross-entropy with squentropy | fine-tuned BERT · MRPC | accuracy: 82.1 -> 84.0 | Improved |
| Replace cross-entropy with squentropy | ResNet-18 · CIFAR-10 | ECE: 10.0 -> 8.9 | Improved |
| Replace cross-entropy with squentropy | Wide ResNet · CIFAR-100 | accuracy: 76.7 -> 77.5 | Improved |
| Replace cross-entropy with squentropy | Transformer-XL · text8 | accuracy: 72.8 -> 71.5 | Hurt |
| Replace cross-entropy with squentropy | EfficientNet · ImageNet | accuracy: 77.0 -> 76.4 | Hurt |
| Replace cross-entropy with squentropy | models across NLP, speech, and vision tasks · 34 classification tasks | classification accuracy or error rate: 24 out of 34 | Mixed |
| Replace cross-entropy with squentropy | models across NLP, speech, and vision tasks · 32 tasks with calibration results | ECE: 26 out of 32 | Improved |
| Replace cross-entropy with squentropy | 3-layer fully connected network · 121 tabular datasets | test accuracy: 94 out of 121 | Mixed |
| Replace cross-entropy with squentropy | 3-layer fully connected network · 121 tabular datasets | ECE: 83 | Improved |
| Use squentropy instead of cross-entropy | models across NLP, speech, and vision tasks · initialization robustness across reported tasks | standard deviation of test performance | Improved |
| Use squentropy instead of cross-entropy | ResNet-18 · STL-10 | accuracy: 93.7 -> 93.0 | Hurt |
Closest papers by meaning
- Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification (NeurIPS 2026 · similarity 0.85)
- EnsLoss: Stochastic Calibrated Loss Ensembles for Preventing Overfitting in Classification (ICML 2025 · similarity 0.84)
- BCE vs. CE in Deep Feature Learning (ICML 2025 · similarity 0.84)
- Neural Collapse for Cross-entropy Class-Imbalanced Learning with Unconstrained ReLU Features Model (ICML 2024 · similarity 0.84)
- Engineering the Neural Collapse Geometry of Supervised-Contrastive Loss (Student Abstract) (AAAI 2024 · similarity 0.84)
- Improved Balanced Classification with Theoretically Grounded Loss Functions (NeurIPS 2025 · similarity 0.83)
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.
