Research map / Label-efficient and robust learning / Recognition and long-tail
Maximum Likelihood with Bias-Corrected Calibration is Hard-To-Beat at Label Shift Adaptation
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 |
|---|---|---|---|
| Combine EM with BCTS rather than leaving predictions uncalibrated (ablation) | Label shift estimation on CIFAR10 · CIFAR10 dirichlet shift, α=0.1, n=2000 | shift-weight MSE: 0.01049 -> 0.00037 | Helped |
| Use EM with bias-corrected calibration instead of BBSL-soft or RLLS-soft | CIFAR10 label shift estimation · CIFAR10 dirichlet shift, α=0.1, n=2000 | shift-weight MSE: 0.00309 -> 0.00037; 0.00316 -> 0.00037 | Helped |
| Add class-specific bias correction to calibration (ablation) | EM label shift adaptation · Across CIFAR10, CIFAR100, MNIST, and diabetic retinopathy shifts | shift-weight MSE: CIFAR10: 0.00944 -> 0.00037 | Helped |
| Use temperature scaling instead of bias-corrected calibration (ablation) | EM label shift adaptation · CIFAR10, CIFAR100, and diabetic retinopathy accuracy tests | Δ% accuracy: CIFAR10: 5.6 vs 5.825; CIFAR100: 24.479 vs 24.829; diabetic retinopathy: 1.6 vs 3.6 | Hurt |
| Compute source priors from predicted probabilities rather than labels when predictions retain bias (ablation) | EM on diabetic retinopathy · Diabetic retinopathy, ρ=0.5, n=500, no calibration | Δ% accuracy: -3.4 -> 2.0 | Helped |
| Compute source priors from predicted probabilities rather than labels with temperature scaling (ablation) | EM on diabetic retinopathy · Diabetic retinopathy, ρ=0.5, n=500, TS | Δ% accuracy: -64.5 -> 2.2 | Helped |
| Compute source priors from predicted probabilities rather than labels with bias-corrected calibration (ablation) | EM on diabetic retinopathy · Diabetic retinopathy, ρ=0.5, n=500, BCTS | Δ% accuracy: 3.8 vs 3.8 | No clear effect |
| Use vector scaling rather than temperature scaling for calibration (ablation) | CIFAR10 calibration quality · CIFAR10 unshifted test set, n=2000 | NLL: 0.292 -> 0.276 | Helped |
| Use BCTS rather than temperature scaling for calibration (ablation) | CIFAR10 calibration quality · CIFAR10 unshifted test set, n=2000 | NLL: 0.292 -> 0.274 | Helped |
| Use calibration methods with class-specific bias parameters (ablation) | EM label shift adaptation · Across tested datasets and shifts | Δ% accuracy: CIFAR10: 5.6 -> 5.825 | Helped |
Closest papers by meaning
- LaSCal: Label-Shift Calibration without target labels (NeurIPS 2024 · similarity 0.89)
- Detecting and Correcting for Label Shift with Black Box Predictors (ICML 2018 · similarity 0.89)
- Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition (ICML 2024 · similarity 0.86)
- Hypothesis Testing for Class-Conditional Noise Using Local Maximum Likelihood (AAAI 2024 · similarity 0.85)
- Graph–Smoothed Bayesian Black-Box Shift Estimator and Its Information Geometry (NeurIPS 2025 · similarity 0.85)
- Bridging Multicalibration and Out-of-distribution Generalization Beyond Covariate Shift (NeurIPS 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.