Research map / Optimization / Optimizers
Stochastic Gradient Descent as Approximate Bayesian Inference
Stephan Mandt, Matthew D. Hoffman, David M. Blei
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 |
|---|---|---|---|
| add full-rank preconditioning to constant SGD (ablation) | constant SGD · Wine | KL divergence to posterior: 18.7 -> 0.7 | Improved |
| add full-rank preconditioning to constant SGD (ablation) | constant SGD · Skin | KL divergence to posterior: 0.471 -> 0.005 | Improved |
| add full-rank preconditioning to constant SGD (ablation) | constant SGD · Protein | KL divergence to posterior: 1000.9 -> 1.8 | Improved |
| add diagonal preconditioning to constant SGD (ablation) | constant SGD · Wine, Skin, Protein | KL divergence to posterior: 18.7 -> 14.0 (Wine), 0.471 -> 0.921 (Skin), 1000.9 -> 678.4 (Protein) | Mixed |
| replace SGLD with constant SGD-f as approximate posterior sampler | SGLD · Wine, Skin, Protein | KL divergence to posterior: 2.9 -> 0.7 (Wine), 0.905 -> 0.005 (Skin), 4.5 -> 1.8 (Protein) | Improved |
| replace BBVI with constant SGD-f as posterior approximation | BBVI · Wine, Skin, Protein | KL divergence to posterior: 44.7 -> 0.7 (Wine), 5.74 -> 0.005 (Skin), 478.1 -> 1.8 (Protein) | Improved |
| use constant SGD-f instead of SGFS-f | SGFS-f · Wine, Skin, Protein | KL divergence to posterior: 0.8 -> 0.7 (Wine), 0.005 -> 0.005 (Skin), 1.3 -> 1.8 (Protein) | No clear effect |
| use IASG with constant learning rate instead of SGLD with decreasing rate | SGLD · linear regression posterior variance estimation | standard error of marginal variance estimates | Improved |
| use IASG instead of NUTS and SGLD under random initialization | NUTS and SGLD · linear regression posterior covariance, random init | posterior covariance structure recovery | Improved |
| use constant SGD hyperparameter selection instead of BBVI for L2 strength | BBVI · MNIST and cover type | validation loss | No clear effect |
Closest papers by meaning
- A Variational Analysis of Stochastic Gradient Algorithms (ICML 2016 · similarity 0.95)
- Structured Stochastic Gradient MCMC (ICML 2022 · similarity 0.92)
- Stochastic Gradient Descent under Markovian Sampling Schemes (ICML 2023 · similarity 0.91)
- Gaussian Approximation and Concentration of Constant Learning-Rate Stochastic Gradient Descent (NeurIPS 2025 · similarity 0.90)
- Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo (ICLR 2025 · similarity 0.89)
- Nonasymptotic Analysis of Stochastic Gradient Descent with the Richardson–Romberg Extrapolation (ICLR 2025 · similarity 0.89)
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.
