Research map / Label-efficient and robust learning / Recognition and long-tail
On the consistency of top-k surrogate losses
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 |
|---|---|---|---|
| Use ψ5, the hinge loss with the k+1 score margin | Top-k classification · Top-k classification, unrestricted predictors | top-k calibration | Helped |
| Use ψ1, ψ2, ψ3, or ψ4 instead of a calibrated surrogate | Top-k classification · Top-k classification, unrestricted predictors | top-k calibration | Hurt |
| Use cross entropy with unrestricted scoring functions | Top-k classification · Top-k classification, unrestricted predictors | top-k calibration | Helped |
| Restrict cross-entropy predictors to linear functions when k=2 | Cross entropy · Top-k separable linear predictors, d≥3, M≥3 | top-k consistency | Hurt |
| Use ψ1 or ψ5 with linear predictors on top-k separable data | Top-k classification · Top-k separable linear predictors | top-k consistency | Helped |
| Use ψ5 rather than ψ1–ψ4 | Hinge surrogate training · Synthetic hinge experiment (training set) | Top-2 accuracy: 0.2671 -> 0.2941 | Helped |
| Use ψ5⁵ rather than entropy | Linear predictors on overlapping synthetic distributions · Synthetic overlapping distributions, N=50 | Top-5 accuracy: 0.755 -> 0.880 | Helped |
| Use ψ5⁵ rather than entropy | Linear predictors on overlapping synthetic distributions · Synthetic overlapping distributions, N=10 | accuracy: 0.212 -> 0.117 | Hurt |
| Set ψ5's loss parameter to k=4 rather than k=5 | ψ5 · Synthetic overlapping distributions, N=100 | Top-5 accuracy: 0.884 -> 0.868 | No clear effect |
| Use entropy rather than ψ5⁵ | Linear predictors on separated synthetic distributions · Synthetic separated distributions, N=100 | Top-5 accuracy: 0.695 -> 0.888 | Helped |
| Use ψ5⁵ rather than entropy | Linear predictors on separated synthetic distributions · Synthetic separated distributions, N=10 | Top-5 accuracy: 0.932 -> 0.874 | No clear effect |
| Use an inverse top-k preserving map with a Bregman divergence | Bregman-divergence surrogate · Top-k classification | top-k calibration | Helped |
Closest papers by meaning
- Cardinality-Aware Set Prediction and Top-$k$ Classification (NeurIPS 2024 · similarity 0.86)
- Consistency Conditions for Differentiable Surrogate Losses (NeurIPS 2025 · similarity 0.85)
- The Minimax Rate of Second-Order Calibration (NeurIPS 2026 · similarity 0.85)
- Optimal and Provable Calibration in High-Dimensional Binary Classification: Angular Calibration and Platt Scaling (NeurIPS 2025 · similarity 0.85)
- Trading off Consistency and Dimensionality of Convex Surrogates for Multiclass Classification (NeurIPS 2024 · similarity 0.85)
- Realizable $H$-Consistent and Bayes-Consistent Loss Functions for Learning to Defer (NeurIPS 2024 · similarity 0.85)
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.