Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Disentangling the Roles of Curation, Data-Augmentation and the Prior in the Cold Posterior Effect
Lorenzo Noci, Kevin Roth, Gregor Bachmann, Sebastian Nowozin, Thomas Hofmann
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 |
|---|---|---|---|
| Turn off data augmentation (ablation) | SG-MCMC BNNs on SVHN and CIFAR-10 · SVHN and CIFAR-10 | test cross-entropy | No clear effect |
| Curate only the training set (ablation) | SG-MCMC BNN on synthetically curated SVHN · SVHN, synthetic curation | CPE presence | No clear effect |
| Curate both training and test sets (ablation) | SG-MCMC BNN on synthetically curated SVHN · SVHN, synthetic curation | CPE presence | Improved |
| Add data augmentation (ablation) | SG-MCMC BNN · SVHN | test cross-entropy | Improved |
| Add data augmentation (ablation) | SG-MCMC BNN · CIFAR-10 | test cross-entropy | Improved |
| Train on smaller random subsets (ablation) | SG-MCMC BNN · SVHN | CPE ratio | Mixed |
| Train on smaller random subsets (ablation) | SG-MCMC BNN · CIFAR-10 | CPE ratio | No clear effect |
| Train on smaller random subsets (ablation) | 1-hidden-layer MLP · Toy 2D classification | CPE ratio | Improved |
| Overweight the likelihood (ablation) | ResNet-20 on CIFAR-10H · CIFAR-10H | CPE presence | Improved |
| Use fewer sampled annotator labels per image (ablation) | ResNet-20 on CIFAR-10H · CIFAR-10H, subsampled labels | CPE presence | Mixed |
| Increase prior influence by reducing dataset size (ablation) | Bayesian neural networks · Random subsampling experiments | CPE strength | Mixed |
Closest papers by meaning
- How Tempering Fixes Data Augmentation in Bayesian Neural Networks (ICML 2022 · similarity 0.90)
- How Good is the Bayes Posterior in Deep Neural Networks Really? (ICML 2020 · similarity 0.89)
- Recursive PAC-Bayes: A Frequentist Approach to Sequential Prior Updates with No Information Loss (NeurIPS 2024 · similarity 0.84)
- A Symmetry-Aware Exploration of Bayesian Neural Network Posteriors (ICLR 2024 · similarity 0.84)
- Scalable Bayesian Learning with posteriors (ICLR 2025 · similarity 0.83)
- How Diffusion Models Memorize (NeurIPS 2026 · similarity 0.83)
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.
