Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Collapsed Inference for Bayesian Deep Learning
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 |
|---|---|---|---|
| Use CIBER collapsed samples and WMI marginalization instead of SGLD | VGG-16 · CIFAR-10 | NLL: 0.2001 $\pm$ 0.0059 -> 0.1927 $\pm$ 0.0029 | Helped |
| Use CIBER collapsed samples and WMI marginalization instead of SGLD | VGG-16 · CIFAR-100 | NLL: 0.9699 $\pm$ 0.0057 -> 0.9193 $\pm$ 0.0027 | Helped |
| Use CIBER instead of SGLD | VGG-16 · CIFAR-10 | ECE: 0.0082 $\pm$ 0.0012 -> 0.0130 $\pm$ 0.0011 | Hurt |
| Use CIBER instead of SGLD | VGG-16 · CIFAR-100 | ECE: 0.0424 $\pm$ 0.0029 -> 0.0168 $\pm$ 0.0025 | Helped |
| Use CIBER instead of the strongest reported log-likelihood baseline | Bayesian regression model · Yacht | test log likelihood: -0.225 $\pm$ 0.400 -> -0.678 $\pm$ 0.301 | Hurt |
| Replace Monte Carlo point estimates with uniform collapsed posterior integration | CIBER · Bayesian linear regression | KL divergence: 0.130 -> 0.069 | Helped |
| Use CIBER instead of SWAG | VGG-16 · CIFAR-10 to STL-10 | NLL: 1.3425 $\pm$ 0.0015 -> 0.9869 $\pm$ 0.0102 | Helped |
| Choose the last layer rather than three weights at the second-to-last layer as the collapsed set (ablation) | CIBER · Concrete | test log likelihood: -2.975 $\pm$ 0.102 -> -2.959 $\pm$ 0.109 | Helped |
| Choose the last layer rather than three weights at the second-to-last layer as the collapsed set (ablation) | CIBER · Naval | test log likelihood: 7.276 $\pm$ 0.532 -> 7.482 $\pm$ 0.188 | Helped |
| Choose the last layer rather than three weights at the second-to-last layer as the collapsed set (ablation) | CIBER · KeggD | test log likelihood: 1.245 $\pm$ 0.090 -> 1.178 $\pm$ 0.088 | Hurt |
| Choose the last layer rather than three weights at the second-to-last layer as the collapsed set (ablation) | CIBER · KeggU | test log likelihood: 1.125 $\pm$ 0.269 -> 0.964 $\pm$ 0.231 | Hurt |
| Choose the last layer rather than three weights at the second-to-last layer as the collapsed set (ablation) | CIBER · Pol | test log likelihood: 2.555 $\pm$ 0.115 -> 2.506 $\pm$ 0.150 | Hurt |
Closest papers by meaning
- Collapsed Variational Bounds for Bayesian Neural Networks (NeurIPS 2021 · similarity 0.89)
- On the Expressiveness of Approximate Inference in Bayesian Neural Networks (NeurIPS 2020 · similarity 0.89)
- A Study of Bayesian Neural Network Surrogates for Bayesian Optimization (ICLR 2024 · similarity 0.88)
- Training Bayesian Neural Networks with Sparse Subspace Variational Inference (ICLR 2024 · similarity 0.87)
- A Symmetry-Aware Exploration of Bayesian Neural Network Posteriors (ICLR 2024 · similarity 0.87)
- Connecting the Dots: Is Mode-Connectedness the Key to Feasible Sample-Based Inference in Bayesian Neural Networks? (ICML 2024 · similarity 0.86)
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.