Research map / Learning theory and algorithms / Prediction and decision losses
Top-Ambiguity Samples Matter: Understanding Why Deep Ensemble Works in Selective Classification
Qiang Ding, Yixuan Cao, Ping Luo
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 |
|---|---|---|---|
| ensemble 5 members vs single member (Deep Ensemble) | Deep Ensemble · CIFAR-10 | AURC/10^-4: 66.1±5.6 -> 51.2±2.0 | Improved |
| ensemble vs single member (Deep Ensemble) | Deep Ensemble · CIFAR-100 | AURC/10^-4: 793.1±9.0 -> 672.5±1.9 | Improved |
| ensemble vs single member (Gambler) | Gambler · CIFAR-100 | AURC/10^-4: 930.6±10.2 -> 857.1±4.7 | Improved |
| ensemble vs single member (Gambler) | Gambler · SVHN | AURC/10^-4: 44.2±2.5 -> 37.0±0.8 | Improved |
| ensemble only on top-50% ambiguity samples vs lowest-50% (ablation) | Deep Ensemble · CIFAR/text datasets | selective risk | Improved |
| ensemble only on low-ambiguity samples (ablation) | Deep Ensemble · CIFAR/text datasets | selective risk | No clear effect |
| increase number of ensemble members (ablation) | Deep Ensemble · CIFAR-10 | AURC | Improved |
| ensemble of 4 VGG-16 vs a single Big VGG-16 of similar size (ablation) | Deep Ensemble · CIFAR-10 | AURC/10^-4: 89.2 -> 49.3 | Improved |
| Deep Ensemble vs SAT ensemble on noisy-label dataset | Deep Ensemble · SVHN | AURC/10^-4: 32.4±0.9 -> 37.1±0.3 | Hurt |
| remove label noise from SVHN (ablation) | Deep Ensemble · SVHN | AURC | Improved |
| Reg-curr ensemble vs Deep Ensemble | Reg-curr · MNLI | AURC/10^-4: 465.2 ±3.8 -> 451.9 ±2.1 | Improved |
| Gambler ensemble vs Deep Ensemble | Gambler · MRPC | AURC/10^-4: 561.3±55.6 -> 650.1±24.8 | Hurt |
Closest papers by meaning
- Theoretical Limitations of Ensembles in the Age of Overparameterization (ICML 2025 · similarity 0.85)
- What Does It Take to Build a Performant Selective Classifier? (NeurIPS 2025 · similarity 0.85)
- No Free Lunch from Random Feature Ensembles: Scaling Laws and Near-Optimality Conditions (ICML 2025 · similarity 0.85)
- Don't Always Pick the Highest-Performing Model: An Information Theoretic View of LLM Ensemble Selection (NeurIPS 2026 · similarity 0.84)
- Sharpness-diversity tradeoff: improving flat ensembles with SharpBalance (NeurIPS 2024 · similarity 0.84)
- United We Stand: Using Epoch-Wise Agreement of Ensembles to Combat Overfit (AAAI 2024 · 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.
