Research map / Label-efficient and robust learning / Recognition and long-tail
Combining Human Predictions with Model Probabilities via Confusion Matrices and Calibration
Gavin Kerrigan, Padhraic Smyth, Mark Steyvers
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 |
|---|---|---|---|
| Combine calibrated model probabilities with human labels | ResNet-164 · CIFAR-10H | error rate (%): 6.10 ± 0.38 -> 2.78 ± 0.15 | Improved |
| Combine calibrated model probabilities with human labels | VGG-19 · ImageNet-16H, noise 80 | error rate (%): 11.10 ± 0.60 -> 6.03 ± 0.54 | Improved |
| Add temperature scaling to the human-model combination (ablation) | P+L on ResNet-164 · CIFAR-10H | error rate (%): 3.05 ± 0.23 -> 2.78 ± 0.15 | Improved |
| Use ensemble temperature scaling instead of temperature scaling (ablation) | P+L on ResNet-164 · CIFAR-10H | error rate (%): 2.78 ± 0.15 -> 2.82 ± 0.23 | No clear effect |
| Use IMax binning instead of temperature scaling (ablation) | P+L on ResNet-164 · CIFAR-10H | error rate (%): 2.78 ± 0.15 -> 2.85 ± 0.23 | No clear effect |
| Discard model probabilities and combine hard labels with confusion matrices (ablation) | ResNet-164 plus human · CIFAR-10H | error rate (%): 2.78 ± 0.15 -> 4.71 ± 0.37 | Hurt |
| Fit the P+L combination with 10 rather than 5000 labeled datapoints (ablation) | ResNet-164 P+L · CIFAR-10H | ECE (10^-2): 1.25 ± 0.18 -> 1.39 ± 0.18 | No clear effect |
| Add post-hoc recalibration to the combined predictions (ablation) | ResNet-164 P+L · CIFAR-10H | ECE (10^-2): 1.39 ± 0.18 -> 0.84 ± 0.20 | Improved |
| Replace temperature scaling with IMax binning (ablation) | P+L on VGG-19 · ImageNet-16H, noise 125 | error rate (%): 19.63 ± 0.70 -> 19.45 ± 0.62 | Improved |
| Use MAP rather than maximum-likelihood parameter estimation (ablation) | Human-model combination · CIFAR-10H and ImageNet-16H | data efficiency | Improved |
| Fit the combination without ground-truth labels using EM (ablation) | P+L-EM · CIFAR-10H and ImageNet-16H | error rate | Mixed |
Closest papers by meaning
- Human and AI Perceptual Differences in Image Classification Errors (AAAI 2025 · similarity 0.86)
- Bayesian Inference for Correlated Human Experts and Classifiers (ICML 2025 · similarity 0.85)
- L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement (ICLR 2025 · similarity 0.85)
- Towards Human-AI Complementarity with Prediction Sets (NeurIPS 2024 · similarity 0.84)
- Quantifying Uncertainty in Error Consistency: Towards Reliable Behavioral Comparison of Classifiers (NeurIPS 2025 · similarity 0.84)
- Intriguing Properties of Generative Classifiers (ICLR 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.
