Research map / Learning theory and algorithms / Prediction and decision losses
In Defense of Softmax Parametrization for Calibrated and Consistent Learning to Defer
Yuzhou Cao, Hussein Mozannar, Lei Feng, Hongxin Wei, Bo An
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 |
|---|---|---|---|
| Use asymmetric softmax surrogate (A-SM) instead of symmetric softmax (S-SM) | L2D surrogate · CIFAR100 expert acc 94%, k=20 | coverage (%): 77.72 -> 98.16 | Improved |
| Use A-SM instead of S-SM with a high-accuracy expert | L2D surrogate · CIFAR100 expert acc 94%, k=60 | error (%): 18.69 -> 19.30 | No clear effect |
| Use A-SM instead of S-SM with a 75% accuracy expert | L2D surrogate · CIFAR100 expert acc 75%, k=60 | error (%): 24.05 -> 22.48 | Improved |
| Use A-SM instead of S-SM | L2D surrogate · CIFAR100 expert acc 94%, k=60 | ECE (%): 11.10 -> 5.96 | Improved |
| Use A-SM instead of A-OvA | L2D surrogate · CIFAR100 expert acc 94%, k=40 | coverage (%): 80.46 -> 92.20 | Improved |
| Use A-SM instead of A-OvA | L2D surrogate · CIFAR100 expert acc 94%, k=40 | ECE (%): 6.81 -> 6.58 | No clear effect |
| Use bounded asymmetric OvA instead of symmetric OvA (S-OvA) | L2D surrogate · CIFAR100 expert acc 94%, k=20 | coverage (%): 85.40 -> 90.40 | Improved |
| Use A-SM instead of S-SM under 10% deferral budget | L2D surrogate · CIFAR100 expert acc 94%, k=40 | budgeted error (%): 46.69 -> 22.17 | Improved |
| Use A-SM instead of A-OvA under 10% deferral budget | L2D surrogate · HateSpeech | budgeted error (%): 16.525 -> 15.07 | Improved |
| Use A-SM instead of S-SM | L2D surrogate · HateSpeech | error (%): 8.65 -> 8.06 | Improved |
| Use A-SM instead of A-OvA | L2D surrogate · ImageNet-16H noise 095 | error (%): 14.17 -> 12.59 | Improved |
| Use A-SM instead of A-OvA | L2D surrogate · CIFAR10H | ECE (%): 0.95 -> 0.94 | No clear effect |
Closest papers by meaning
- Exploiting Human-AI Dependence for Learning to Defer (ICML 2024 · similarity 0.87)
- Realizable $H$-Consistent and Bayes-Consistent Loss Functions for Learning to Defer (NeurIPS 2024 · similarity 0.86)
- Regression with Multi-Expert Deferral (ICML 2024 · similarity 0.86)
- Soft ascent-descent as a stable and flexible alternative to flooding (NeurIPS 2024 · similarity 0.84)
- Mastering Multiple-Expert Routing: Realizable $H$-Consistency and Strong Guarantees for Learning to Defer (ICML 2025 · similarity 0.84)
- Density-Ratio Losses for Post-Hoc Learning to Defer (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.
