Research map / Label-efficient and robust learning / Recognition and long-tail
DICE: Diversity in Deep Ensembles via Conditional Redundancy Adversarial Estimation
Alexandre Rame, Matthieu Cord
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 |
|---|---|---|---|
| Add conditional redundancy loss to CEB (ablation) | DICE ensemble · CIFAR-100, ResNet-32 4-branch | Top-1 accuracy (%): 76.98 -> 77.51 | Improved |
| Use conditional rather than non-conditional redundancy (ablation) | VCEB ensemble · CIFAR-100, ResNet-32 4-branch | Top-1 accuracy (%): 77.30 -> 77.51 | Improved |
| Use VCEB instead of VIB (ablation) | Ensemble without redundancy · CIFAR-100, ResNet-32 4-branch | Top-1 accuracy (%): 76.93 -> 76.98 | No clear effect |
| Add non-conditional redundancy to VIB (ablation) | IBR ensemble · CIFAR-100, ResNet-32 4-branch | Top-1 accuracy (%): 76.93 -> 77.25 | Improved |
| Add conditional redundancy loss (ablation) | DICE ensemble · CIFAR-100, ResNet-110 4-branch | Top-1 accuracy (%): 81.17 -> 81.93 | Improved |
| Add conditional redundancy loss (ablation) | DICE ensemble · CIFAR-100, ResNet-32 4-branch | NLL (10^-1), after temperature scaling: 8.11 -> 7.98 | Improved |
| Add conditional redundancy loss (ablation) | DICE ensemble · CIFAR-100, ResNet-32 4-branch | ECE (10^-2), before temperature scaling: 5.76 -> 4.05 | Improved |
| Scale logits with the discriminator's 1-w confidence | DICE×w · CIFAR-100, TinyImageNet (crop) OOD | AUROC: 74.7 -> 77.2 | Improved |
| Increase the diversity coefficient to 0.6 (ablation) | DICE · CIFAR-100, 4-branch ResNet-32 | Individual accuracy | Hurt |
| Set the diversity coefficient to -0.025 (ablation) | DICE · CIFAR-100, 4-branch ResNet-32 | Ensemble diversity | Hurt |
| Use adversarial redundancy loss without compression (ablation) | Redundancy-only ensemble · CIFAR-100 | Training stability | Hurt |
Closest papers by meaning
- Improving Adversarial Robustness via Promoting Ensemble Diversity (ICML 2019 · similarity 0.90)
- On the Diversity of Adversarial Ensemble Learning (ICML 2025 · similarity 0.89)
- DIBS: Diversity Inducing Information Bottleneck in Model Ensembles (AAAI 2021 · similarity 0.88)
- Diversity Matters When Learning From Ensembles (NeurIPS 2021 · similarity 0.88)
- Sharpness-diversity tradeoff: improving flat ensembles with SharpBalance (NeurIPS 2024 · similarity 0.86)
- The Curse of Diversity in Ensemble-Based Exploration (ICLR 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.
