Research map / Probabilistic and scientific modeling / Variational and generative sampling
Amortized Bethe Free Energy Minimization for Learning MRFs
Sam Wiseman, Yoon Kim
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 amortized BFE minimization instead of Loopy BP | Ising model learning · Ising models, n=15 | held out NLL: 59.79 -> 54.91 | Improved |
| Use amortized BFE minimization instead of exact inference | Ising model learning · Ising models, n=15 | held out NLL: 52.24 -> 54.91 | Hurt |
| Use the inference network for marginal estimation instead of Loopy BP | Ising model inference · Ising models, n=15 | marginal correlation: 0.942 -> 0.981 | Improved |
| Use the inference network for marginal estimation instead of Loopy BP | Ising model inference · Ising models, n=15 | Mean L1 distance: 0.065 -> 0.040 | Improved |
| Use amortized BFE minimization instead of Loopy BP | RBM learning · UCI digits | held out average NLL: 25.47 -> 23.43 | Improved |
| Use amortized BFE minimization instead of PCD | RBM learning · UCI digits | held out average NLL: 21.24 -> 23.43 | Hurt |
| Use amortized BFE minimization instead of Loopy BP | Undirected 3rd Order HMM learning · Penn Treebank | average NLL: 108.74 -> 115.86 | Hurt |
| Use an undirected pairwise-factor HMM and amortized BFE rather than the directed model with a mean-field VAE | 3rd Order HMM learning · Penn Treebank | average NLL: 119.27 -> 115.86 | Improved |
| Use an undirected pairwise-factor HMM and amortized BFE rather than the directed model with a 1st Order HMM VAE | 3rd Order HMM learning · Penn Treebank | average NLL: 118.35 -> 115.86 | Improved |
| Use amortized BFE minimization rather than Loopy BP | Undirected 3rd Order HMM learning · Penn Treebank | seconds/epoch: 247 -> 70 | Improved |
| Replace a Transformer inference network with a BLSTM (ablation) | RBM and HMM inference networks · RBM and Penn Treebank HMM experiments | held out performance | Improved |
| Use KL divergence rather than L2 distance for the consistency penalty (ablation) | RBM and HMM inference networks · RBM and Penn Treebank HMM experiments | held out performance | Improved |
Closest papers by meaning
- Adaptive Variational Inference: Beyond Bethe, Tree-Reweighted, and Convex Free Energies (TPAMI 2026 · similarity 0.86)
- Iterative Amortized Inference (ICML 2018 · similarity 0.86)
- Efficient semidefinite-programming-based inference for binary and multi-class MRFs (NeurIPS 2020 · similarity 0.85)
- Energy-Inspired Models: Learning with Sampler-Induced Distributions (NeurIPS 2019 · similarity 0.85)
- Learning Energy-Based Model with Variational Auto-Encoder as Amortized Sampler (AAAI 2021 · similarity 0.85)
- Meta-Learning with Shared Amortized Variational Inference (ICML 2020 · similarity 0.85)
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.
