Research map / Label-efficient and robust learning / Recognition and long-tail
Elodi: Ensemble Logit Difference Inhibition for Positive-Congruent Training
Yue Zhao, Yantao Shen, Yuanjun Xiong, Shuo Yang, Wei Xia, Zhuowen Tu, Bernt Schiele, Stefano Soatto
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 |
|---|---|---|---|
| distill from a homogeneous 8x ensemble with LDI loss (Elodi K=10) instead of no treatment | RN-18 to RN-50 update · ImageNet | NFR (%): 4.30 -> 2.11 | Improved |
| use top-K=10 classes in LDI instead of all 1000 (ablation) | Elodi · ImageNet | NFR (%): 2.18 -> 2.11 | No clear effect |
| Elodi instead of targeted FD-KL | RN-18 to RN-50 update · ImageNet | NFR (%): 2.90 -> 2.11 | Improved |
| Elodi instead of dropout as implicit ensemble | RN-18 to RN-50 update · ImageNet | Rel-NFR (%): 21.92 -> 13.23 | Improved |
| Elodi instead of Focal Distillation, data growth with more classes | RN-18 to RN-50 · ImageNet (increasing classes) | NFR (%): 5.45 -> 4.10 | Improved |
| Elodi instead of Focal Distillation, more samples per class | RN-18 to RN-50 · ImageNet (increasing samples/class) | NFR (%): 2.65 -> 2.21 | Improved |
| Elodi on AG News vs no treatment | BERT-base update · AG News | NFR (%): 0.934 -> 0.289 | Improved |
| Elodi on chain of three models vs no treatment | RN-18, RN-50, RN-101 · ImageNet | pairwise NFR (%): 4.28 -> 2.04 | Improved |
| add LDI loss toward old model when old model was not trained with Elodi (both updates) | Elodi · ImageNet RN-18 to RN-50 | NFR (%): 2.98 -> 2.85 | Improved |
| use homogeneous ensemble instead of all-diff-weak heterogeneous ensemble (ablation) | Elodi guiding ensemble · ImageNet RN-18 to RN-50 | NFR (%): 2.25 -> 2.18 | No clear effect |
| calibrate heterogeneous ensembles with temperature scaling (ablation) | Elodi all-diff ensemble · ImageNet RN-18 to RN-50 | NFR (%): 2.25 -> 2.35 | No clear effect |
| guide new model with ensemble of new architecture instead of old RN-18 architecture (ablation) | Elodi · ImageNet RN-18 to RN-50 | RN-50 error rate (%): 26.82 -> 23.15 | Improved |
Later papers that compared with this method
1 later accepted paper reports ELODI as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Mitigating Negative Flips via Margin Preserving Training (AAAI 2026 · better on 7 of 7 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Multiplicative Logit Adjustment Approximates Neural-Collapse-Aware Decision Boundary Adjustment (ICLR 2025 · similarity 0.83)
- EnsLoss: Stochastic Calibrated Loss Ensembles for Preventing Overfitting in Classification (ICML 2025 · similarity 0.82)
- Subsampled Ensemble Can Improve Generalization Tail Exponentially (NeurIPS 2025 · similarity 0.82)
- Split-Ensemble: Efficient OOD-aware Ensemble via Task and Model Splitting (ICML 2024 · similarity 0.82)
- Invertible Logits Transformation for Accuracy-Preserving Post-Hoc Uncertainty Calibration (NeurIPS 2026 · similarity 0.82)
- Spurious Feature Diversification Improves Out-of-distribution Generalization (ICLR 2024 · similarity 0.82)
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.
