Research map / Label-efficient and robust learning / Test-time adaptation
Distributionally Robust Post-hoc Classifiers under Prior Shifts
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 |
|---|---|---|---|
| Post-hoc DROPS scaling (delta=0.9) instead of training-time logit adjustment or other methods | PreAct-ResNet 18 on class-imbalanced CIFAR-10/100 · CIFAR-10 rho=100, delta=1.0-worst accuracy | delta=1.0-worst case accuracy: +10.92 over Logit-Adj (post-hoc) (75.46 vs 64.54) | Helped |
| Post-hoc DROPS scaling instead of Logit-Adj post-hoc | CIFAR-100 class-imbalanced · CIFAR-100 rho=100 averaged accuracy | averaged class accuracy: DROPS 43.20 vs Logit-Adj (post-hoc) 43.68 | Mixed |
| Post-hoc DROPS scaling on ERM model with validation group labels only | Waterbirds · Waterbirds worst group accuracy | worst group accuracy: +6.49 over ERM (77.14 vs 70.65) | Helped |
| Post-hoc DROPS scaling on ERM model | CelebA · CelebA worst group accuracy | worst group accuracy: +34.55 over ERM (79.44 vs 44.89) | Helped |
| Post-hoc DROPS scaling applied on top of DFR last-layer retraining (DROPS*) (ablation) | CelebA and Waterbirds · CelebA worst group accuracy and Waterbirds worst group accuracy | worst group accuracy: CelebA +1.94 over DFR (90.19 vs 88.25); Waterbirds -0.02 vs DFR (91.99 vs 92.01) | Mixed |
| Post-hoc DROPS scaling on ERM for CelebA vs JTT | CelebA · CelebA delta=0.05-Worst accuracy | delta-worst accuracy: DROPS 86.80 vs JTT 83.34 at delta=0.05 (better); on Averaged, JTT 85.04 vs DROPS 89.05 | Mixed |
| Post-hoc DROPS scaling on ERM for CIFAR-100 rho=100 under controlled shift | CIFAR-100 with KL divergence · CIFAR-100 rho=100 averaged accuracy | averaged class accuracy: DROPS 43.20 vs Logit-Adj 43.03 and post-hoc Logit-Adj 43.68 (mixed; trade-off noted by authors) | Hurt |
| Post-hoc DROPS scaling vs Balanced-Softmax training on CIFAR-10 rho=50 averaged accuracy | CIFAR-10 rho=50 · CIFAR-10 rho=50 averaged class accuracy | averaged class accuracy: DROPS 83.12 vs Balanced-Softmax 78.76 (better); CIFAR-10 rho=10 DROPS 89.17 vs Balanced-Softmax 87.18 | Mixed |
Closest papers by meaning
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · similarity 0.89)
- Tolerant Algorithms for Learning with Arbitrary Covariate Shift (NeurIPS 2024 · similarity 0.86)
- Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 2024 · similarity 0.86)
- Structure-informed Risk Minimization for Robust Ensemble Learning (ICML 2025 · similarity 0.86)
- DRoP: Distributionally Robust Data Pruning (ICLR 2025 · similarity 0.86)
- Optimal Classification under Performative Distribution Shift (NeurIPS 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.