Research map / Learning theory and algorithms / Prediction and decision losses
Learning Optimal Conformal Classifiers
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 |
|---|---|---|---|
| Train end-to-end with ConfTr rather than cross-entropy, then use Thr at test time | Classifier · MNIST, F-MNIST, EMNIST, CIFAR10, CIFAR100 | inefficiency: 2.23 -> 2.18; 2.05 -> 1.69; 2.66 -> 2.66; 2.93 -> 2.88; 10.63 -> 10.78 | Mixed |
| Train end-to-end with ConfTr rather than cross-entropy, then use APS at test time | Classifier · MNIST, F-MNIST, EMNIST, CIFAR10, CIFAR100 | inefficiency: 2.50 -> 2.16; 2.36 -> 1.82; 4.23 -> 2.86; 3.30 -> 3.05; 16.62 -> 12.99 | Helped |
| Add classification loss during ConfTr training (ablation) | ConfTr · MNIST, F-MNIST, EMNIST, CIFAR10, CIFAR100 | inefficiency: 2.18 -> 2.11; 1.69 -> 1.67; 2.66 -> 2.49; 2.88 -> 2.84; 10.78 -> 10.44 | Helped |
| Add classification loss during ConfTr training (ablation) | ConfTr · EMNIST | inefficiency: 2.86 -> 2.87 | No clear effect |
| Train with Bel using ThrL instead of cross-entropy | Classifier · MNIST, EMNIST, CIFAR100 | inefficiency: 2.23 -> 2.70; 2.66 -> 3.48; 10.63 -> 10.91 | Hurt |
| Train with Bel using ThrL instead of cross-entropy | Classifier · F-MNIST | inefficiency: 2.05 -> 1.90 | Helped |
| Use CoverTr with a fixed threshold rather than ConfTr's calibrated minibatch threshold (ablation) | Classifier · MNIST | inefficiency: 2.23 -> 6.34 | Hurt |
| Use CoverTr with ThrLP during training (ablation) | Classifier · MNIST | inefficiency: 2.23 -> 2.50 | Hurt |
| Use batch size 100 instead of 500 during ConfTr (ablation) | ConfTr · MNIST | inefficiency: 2.18 -> 8.04 | Hurt |
| Vary the size weight λ from 0.001 to 10 (ablation) | ConfTr without classification loss · MNIST | inefficiency: 2.18; 2.18; 2.18; 2.19; 2.19; 2.19; 2.16 | No clear effect |
| Vary temperature T across tested values (ablation) | ConfTr without classification loss · MNIST | inefficiency: 2.39; 2.23; 2.2; 2.19; 2.18; 2.2; 2.29 | No clear effect |
| Train ConfTr at α=0.005 rather than α=0.01 (ablation) | ConfTr · MNIST | inefficiency: 2.18 -> 2.17 | No clear effect |
Closest papers by meaning
- Class Adaptive Conformal Training (NeurIPS 2026 · similarity 0.88)
- Verifiably Robust Conformal Prediction (NeurIPS 2024 · similarity 0.88)
- Direct Prediction Set Minimization via Bilevel Conformal Classifier Training (ICML 2025 · similarity 0.87)
- On Temperature Scaling and Conformal Prediction of Deep Classifiers (ICML 2025 · similarity 0.85)
- Reliable Uncertainty Quantification in Machine Learning via Conformal Prediction (AAAI 2025 · similarity 0.85)
- Optimal transport-based conformal prediction (ICML 2025 · 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.