Research map / Learning theory and algorithms / Prediction and decision losses
Smooth Loss Functions for Deep Top-k Classification
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 |
|---|---|---|---|
| Replace cross-entropy with L₅,₁ | DenseNet 40-40 · CIFAR-100, noise 0.8 | top-1 accuracy (%): 35.53 -> 55.85 | Helped |
| Replace cross-entropy with L₅,₁ | DenseNet 40-40 · CIFAR-100, noise 0.8 | top-5 accuracy (%): 74.80 -> 79.32 | Helped |
| Replace cross-entropy with L₅,₁ | DenseNet 40-40 · CIFAR-100, noise 0.0 | top-1 accuracy (%): 76.68 -> 69.33 | Hurt |
| Replace cross-entropy with L₅,₁ | DenseNet 40-40 · CIFAR-100, noise 0.0 | top-5 accuracy (%): 94.34 -> 94.29 | No clear effect |
| Replace cross-entropy with L₅,₀.₁ | ResNet-18 · ImageNet val, 5% training data | top-5 accuracy (%): 58.31 -> 60.44 | Helped |
| Replace cross-entropy with L₅,₀.₁ | ResNet-18 · ImageNet val, 100% training data | top-5 accuracy (%): 90.67 -> 90.61 | No clear effect |
| Use a sufficiently high temperature for L₅,τ rather than a low temperature (ablation) | DenseNet 40-12 on CIFAR-100 · CIFAR-100 training | top-5 training error | Helped |
| Use a very small positive temperature for L₅,τ (ablation) | DenseNet 40-12 on CIFAR-100 · CIFAR-100 training | top-5 training error | Hurt |
Closest papers by meaning
- Stochastic smoothing of the top-K calibrated hinge loss for deep imbalanced classification (ICML 2022 · similarity 0.90)
- Loss Functions for Top-k Error: Analysis and Insights (CVPR 2016 · similarity 0.89)
- Weighted Sampling without Replacement for Deep Top-$k$ Classification (ICML 2023 · similarity 0.89)
- Cross-Entropy Loss Functions: Theoretical Analysis and Applications (ICML 2023 · similarity 0.88)
- ST$_k$: A Scalable Module for Solving Top-k Problems (NeurIPS 2024 · similarity 0.85)
- Loss Functions and Operators Generated by f-Divergences (ICML 2025 · 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.