Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
A statistical theory of cold posteriors in deep neural networks
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 |
|---|---|---|---|
| Model dataset curation as unanimous agreement among multiple labellers | Bayesian neural networks on curated image datasets · Toy GP classification | optimal temperature: S ≈ 1/λ* | Helped |
| Use the correct curation likelihood, including known no-consensus inputs | Tempered GP classification · Toy GP classification with curation | optimal temperature | No clear effect |
| Train and test using the single-labeller likelihood while excluding no-consensus points | Tempered GP classification · Toy GP classification with curated data | optimal temperature: S ≈ 1/λ* | Helped |
| Select high-agreement curated examples rather than random uncurated examples | ResNet18 with SGLD · Galaxy Zoo 2 | test log-likelihood | Helped |
| Select high-agreement curated examples rather than random uncurated examples | ResNet18 with SGLD · Galaxy Zoo 2 | test accuracy | Helped |
| Apply stronger tempering to curated rather than uncurated data | ResNet18 with SGLD · Galaxy Zoo 2 | test log-likelihood: factor of ∼ 3; ∼ 20% | Helped |
| Train on approximately 50 human labels per image instead of one label | ResNet18 with SGLD · CIFAR-10H training / CIFAR-10 test | cold-posterior effect: ∼ 50 | Helped |
| Add noise to labels | ResNet18 with SGLD · CIFAR-10 noisy labels | cold-posterior effect | Mixed |
| Use a temperature close to 1 as label noise increases | ResNet18 with SGLD · CIFAR-10 noisy labels | test performance | Helped |
| Use a tempered MAP objective | ResNet18 without batch normalization · CIFAR-10 | test accuracy | Helped |
Closest papers by meaning
- How Good is the Bayes Posterior in Deep Neural Networks Really? (ICML 2020 · similarity 0.91)
- Bayesian Neural Network Priors Revisited (ICLR 2022 · similarity 0.90)
- A Symmetry-Aware Exploration of Bayesian Neural Network Posteriors (ICLR 2024 · similarity 0.89)
- Do Bayesian Neural Networks Actually Behave Like Bayesian Models? (ICML 2025 · similarity 0.88)
- Make Me a BNN: A Simple Strategy for Estimating Bayesian Uncertainty from Pre-trained Models (CVPR 2024 · similarity 0.87)
- Bayesian Low-Rank Learning (Bella): A Practical Approach to Bayesian Neural Networks (AAAI 2025 · similarity 0.87)
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.