Research map / Learning theory and algorithms / Prediction and decision losses
Fixed Points of Belief Propagation—An Analysis via Polynomial Homotopy Continuation
Christian J. Knoll, Dhagash B. Mehta, Tianran Chen, Franz Pernkopf
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 |
|---|---|---|---|
| find all fixed points with NPHC and pick the best one instead of using the BP-converged fixed point | BP marginals · 3x3 grid, uniform factors, all parameters | MSE: 0.197 -> 0.016 | Improved |
| find all fixed points with NPHC and pick the best one instead of the BP result | BP marginals · 3x3 grid, region (II) | MSE: 0.836 -> 0.050 | Improved |
| find all fixed points with NPHC instead of BP | BP marginals · Fully connected 2x2, region (II) | MSE: 0.304 -> 0.004 | Improved |
| combine only stable fixed points weighted by partition function instead of using the Z_B-maximizing fixed point (ablation) | Max-Z_B fixed point · 3x3 grid, region (I) | MSE: 0.076 -> 1.0x10^-9 | Improved |
| combine all fixed points weighted by partition function instead of using the Z_B-maximizing fixed point (ablation) | Max-Z_B fixed point · 3x3 grid, region (I) | MSE: 0.076 -> 9.2x10^-4 | Improved |
| combine only stable fixed points weighted by partition function instead of the Z_B-maximizing fixed point (ablation) | Max-Z_B fixed point · Fully connected 2x2, region (I) | MSE: 0.070 -> 2.0x10^-8 | Improved |
| combine fixed points weighted by Z_B in region (III) where a unique fixed point exists (ablation) | Single fixed point · 3x3 grid, region (III) | MSE: 0.006 -> 0.006 | No clear effect |
| combine stable fixed points without proper Z_B weighting (ablation) | Stable fixed points · Ising grid and fully connected graphs | accuracy | Hurt |
| increase the local field theta | BP / max-Z_B fixed point · Ising grid and fully connected graphs | convergence and accuracy | Improved |
| use BKK bound instead of total degree to bound the number of solutions | NPHC · 3x3 grid, random factors | number of solution paths: 7.2x10^15 -> 608 | Improved |
| use BKK bound instead of total degree | NPHC · Fully connected 2x2 | number of solution paths: 16.8x10^6 -> 120 | Improved |
| use NPHC instead of BP | BP · 3x3 grid, random factors | runtime (s): 3.6x10^-3 -> 13.57 | Hurt |
Closest papers by meaning
- Self-Guided Belief Propagation – a Homotopy Continuation Method (TPAMI 2022 · similarity 0.93)
- The Linearization of Belief Propagation on Pairwise Markov Random Fields (AAAI 2017 · similarity 0.91)
- Synthesis of MCMC and Belief Propagation (NeurIPS 2016 · similarity 0.89)
- Inference in Graphical Models via Semidefinite Programming Hierarchies (NeurIPS 2017 · similarity 0.89)
- Tensor Belief Propagation (ICML 2017 · similarity 0.88)
- Belief Propagation Converges to Gaussian Distributions in Sparsely-Connected Factor Graphs (ICML 2026 · similarity 0.88)
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.
