Research map / Probabilistic and scientific modeling / Variational and generative sampling
Meta-Learning MCMC Proposals
Tongzhou Wang, YI WU, Dave Moore, Stuart Russell
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 |
|---|---|---|---|
| Use learned block proposals instead of single-site Gibbs updates | Binary grid models · UAI 2008 inference competition grid models | marginal probability error over time | Improved |
| Use a neural block proposal instead of exact block Gibbs | Binary grid models · Binary grid models | error over time | Improved |
| Use an approximate neural block proposal instead of exact block Gibbs | Binary grid models · Binary grid models | error over epochs | Hurt |
| Use a learned block proposal instead of single-site Gibbs | Open-universe GMM inference · Open-universe GMMs of various sizes | mixing speed | Improved |
| Use learned block proposals for consecutive tag variables instead of single-site Gibbs | NER tagging · CoNLL-2003 test sentences | F1 | Improved |
| Use learned block proposals for consecutive tag variables instead of single-site Gibbs | NER tagging · CoNLL-2003 test sentences | log likelihood | Improved |
| Mix single-site Gibbs with neural block proposals | NER tagging · CoNLL-2003 test sentences | mixing performance | Improved |
| Increase the proposed tag block size | NER tagging · CoNLL-2003 test sentences | time to mix | Hurt |
| Collapse the assignment variables and apply the Metropolis-Hastings rule to the collapsed model (ablation) | Neural block proposals for open-universe GMMs · GMMs with varying numbers of observations | scalability | Improved |
| Resample assignments jointly with proposed mixture variables and apply the Metropolis-Hastings rule to all proposed values (ablation) | Neural block proposals for open-universe GMMs · GMMs as the number of observations grows | acceptance ratio | Hurt |
| Use a mixed proposal rather than single-site Gibbs alone | NER tagging · CoNLL-2003 test sentences | F1 | Improved |
| Train on one motif instantiation distribution and apply to other parameter distributions | Grid motif proposal · Random binary-valued grid instantiations | conditional approximation divergence | Improved |
Closest papers by meaning
- Neural Relational Inference with Fast Modular Meta-learning (NeurIPS 2019 · similarity 0.86)
- MARS: Meta-learning as Score Matching in the Function Space (ICLR 2023 · similarity 0.86)
- Recasting Gradient-Based Meta-Learning as Hierarchical Bayes (ICLR 2018 · similarity 0.86)
- A Hierarchical Bayesian Model for Few-Shot Meta Learning (ICLR 2024 · similarity 0.85)
- Context-Aware Meta-Learning (ICLR 2024 · similarity 0.84)
- Learning to Explore for Stochastic Gradient MCMC (ICML 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.
