Research map / Learning theory and algorithms / Prediction and decision losses
Robust Bi-Tempered Logistic Loss Based on Bregman Divergences
Ehsan Amid, Manfred K. Warmuth, Rohan Anil, Tomer Koren
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 logistic loss with bi-tempered loss (0.5, 4.0) | CNN · MNIST, 40% label noise | top-1 accuracy: 97.64 -> 98.56 | Improved |
| Replace logistic loss with bi-tempered loss (0.8, 1.2) | ResNet-56 · CIFAR-100, 40% label noise | top-1 accuracy: 53.17 -> 62.55 | Improved |
| Replace logistic loss with bi-tempered loss (0.9,1.05) | ResNet-18 · ImageNet-2012 | top-1 accuracy: 71.333 -> 71.618 | Improved |
| Replace logistic loss with bi-tempered loss (0.9,1.05) | ResNet-50 · ImageNet-2012 | top-1 accuracy: 76.332 -> 76.748 | Improved |
| Use a heavy-tailed tempered softmax (t2 = 4) while retaining KL loss (t1 = 1) (ablation) | Two-layer network on synthetic small-margin label noise · Synthetic binary classification, small-margin label noise | robustness to label noise | Improved |
| Use a bounded loss (t1 = 0.2) with vanilla softmax (t2 = 1) (ablation) | Two-layer network on synthetic large-margin label noise · Synthetic binary classification, large-margin label noise | robustness to label noise | Improved |
| Use both bounded loss (t1 = 0.2) and heavy-tailed probabilities (t2 = 4.0) (ablation) | Two-layer network on synthetic random label noise · Synthetic binary classification, random label noise | robustness to label noise | Improved |
Closest papers by meaning
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.86)
- Loss Functions and Operators Generated by f-Divergences (ICML 2025 · similarity 0.86)
- Bias-variance decompositions: the exclusive privilege of Bregman divergences (JMLR 2025 · similarity 0.86)
- Cross-Entropy Loss Functions: Theoretical Analysis and Applications (ICML 2023 · similarity 0.85)
- Decoupled Kullback-Leibler Divergence Loss (NeurIPS 2024 · similarity 0.85)
- Robust Diffusion Models via Divergence-Induced Weighted Denoising (NeurIPS 2026 · 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.
