Research map / Learning theory and algorithms / Prediction and decision losses
Fast Differentiable Sorting and Ranking
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 the OT soft ranking operator with r_Q or r_E | vanilla CNN with top-k classification loss · CIFAR-100 | training time for 600 epochs: 29 hours -> 21 hours; 29 hours -> 23 hours | Helped |
| Replace OT with the All-pairs soft ranking operator | vanilla CNN with top-k classification loss · CIFAR-100 | training time for 600 epochs: 29 hours -> 16 hours | Helped |
| Use proposed soft ranks instead of OT soft ranks | vanilla CNN with top-k classification loss · CIFAR-10 and CIFAR-100 | accuracy | No clear effect |
| Use the All-pairs formulation rather than cross-entropy loss | vanilla CNN with top-k classification loss · CIFAR-10 and CIFAR-100 | accuracy | Hurt |
| Use the proposed soft top-k loss rather than cross-entropy loss | vanilla CNN · CIFAR-10 and CIFAR-100 | accuracy | Helped |
| Add a quadratic soft rank layer | linear label-ranking model · 21 label-ranking datasets | Spearman's rank correlation coefficient: 15 datasets; 4; 2 datasets | Mixed |
| Remove the soft ranking layer (ablation) | linear label-ranking model · housing | Spearman's rank correlation coefficient: 0.70 ± 0.03 -> 0.77 ± 0.03 | Hurt |
| Replace quadratic regularization with entropic log-KL projection (ablation) | linear label-ranking model · housing | Spearman's rank correlation coefficient: 0.77 ± 0.03 -> 0.78 ± 0.02 | Helped |
| Replace log-KL projection with direct KL projection (ablation) | linear label-ranking model · housing | Spearman's rank correlation coefficient: 0.78 ± 0.02 -> 0.77 ± 0.03 | No clear effect |
| Add the quadratic soft rank layer (ablation) | linear label-ranking model · pendigits | Spearman's rank correlation coefficient: 0.94 ± 0.00 -> 0.96 ± 0.00 | Helped |
| Scale input dimension using the proposed soft ranking operator rather than OT | batch soft-ranking computation · random Gaussian score vectors, n up to 5000 | memory feasibility: n=2000; n=1000 | Helped |
| Use soft least trimmed squares rather than ridge regression as outliers increase | robust regression · LIBSVM datasets | R_2 | Mixed |
Closest papers by meaning
- Differentiable Ranking and Sorting using Optimal Transport (NeurIPS 2019 · similarity 0.92)
- PiRank: Scalable Learning To Rank via Differentiable Sorting (NeurIPS 2021 · similarity 0.88)
- Monotonic Differentiable Sorting Networks (ICLR 2022 · similarity 0.87)
- LapSum - One Method to Differentiate Them All: Ranking, Sorting and Top-k Selection (ICML 2025 · similarity 0.86)
- Differentiable Sorting Networks for Scalable Sorting and Ranking Supervision (ICML 2021 · similarity 0.86)
- Generalized Neural Sorting Networks with Error-Free Differentiable Swap Functions (ICLR 2024 · 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.