Research map / Learning theory and algorithms / Prediction and decision losses
LegendreTron: Uprising Proper Multiclass Loss Learning
Kevin H Lam, Christian Walder, Spiridon Penev, Richard Nock
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 |
|---|---|---|---|
| Learn a proper multiclass canonical loss and probabilities jointly instead of using multinomial logistic regression | LegendreTron · aloi, sector, letter, news20, Sensorless, vowel, segment, vehicle | classification accuracy: aloi: 10.34$\pm$0.42 -> 88.11$\pm$0.03; sector: 8.77$\pm$0.73 -> 89.71$\pm$0.18; letter: 53.37$\pm$0.25 -> 79.82$\pm$0.3 | Improved |
| Learn a proper multiclass canonical loss and probabilities jointly instead of using multinomial logistic regression | LegendreTron · dna | classification accuracy: 92.79$\pm$0.30 -> 94.43$\pm$0.19 | Hurt |
| Learn a proper multiclass canonical loss and probabilities jointly instead of using multinomial logistic regression | LegendreTron · svmguide2 | classification accuracy: 56.01$\pm$1.40 -> 56.01$\pm$1.40 | No clear effect |
| Learn a proper multiclass canonical loss and probabilities jointly instead of using multinomial logistic regression | LegendreTron · wine | classification accuracy: 96.94$\pm$1.14 -> 97.78$\pm$0.59 | No clear effect |
| Learn a proper multiclass canonical loss and probabilities jointly instead of using multinomial logistic regression | LegendreTron · iris | classification accuracy: 86.67$\pm$3.89 -> 83.00$\pm$2.08 | No clear effect |
| Learn a proper multiclass loss while retaining a VGG-5 model | VGGTron versus VGG · MNIST | classification accuracy: 99.40\% -> 99.59\% | Improved |
| Learn a proper multiclass loss while retaining a VGG-5 model | VGGTron versus VGG · Fashion-MNIST | classification accuracy: 92.80\% -> 92.88\% | Improved |
| Learn a proper multiclass loss while retaining a VGG-5 model | VGGTron versus VGG · Kuzushiji-MNIST | classification accuracy: 98.12\% -> 98.26\% | Improved |
| Model all ten class probabilities jointly rather than using ten one-vs-rest binary models | LegendreTron · MNIST-like datasets | classification accuracy | Improved |
| Use the learned inverse canonical link rather than the identity map | LegendreTron · multiclass classification | classification accuracy | Improved |
| Use softmax+ with an identity inverse-link component, removing learned loss adaptation (ablation) | LegendreTron · MNIST | test AUC: 99.9\% -> 99.9\% | No clear effect |
| Use softmax+ with an identity inverse-link component, removing learned loss adaptation (ablation) | LegendreTron · Fashion-MNIST | test AUC: 99.2\% -> 98.5\% | Hurt |
Closest papers by meaning
- Being Properly Improper (ICML 2022 · similarity 0.86)
- Multiclass Loss Geometry Matters for Generalization of Gradient Descent in Separable Classification (NeurIPS 2025 · similarity 0.84)
- Loss Functions and Operators Generated by f-Divergences (ICML 2025 · similarity 0.83)
- Trading off Consistency and Dimensionality of Convex Surrogates for Multiclass Classification (NeurIPS 2024 · similarity 0.83)
- Learning with Fitzpatrick Losses (NeurIPS 2024 · similarity 0.83)
- Automated Loss function Search for Class-imbalanced Node Classification (ICML 2024 · similarity 0.82)
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.
