Research map / Probabilistic and scientific modeling / Variational and generative sampling
Joint Bayesian Inference of Graphical Structure and Parameters with a Single Generative Flow Network
Tristan Deleu, Mizu Nishikawa-Toomey, Jithendaraa Subramanian, Nikolay Malkin, Laurent Charlin, Yoshua Bengio
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 JSP-GFN (diag) instead of best baseline BCD Nets | joint posterior approximation · d=5 linear Gaussian | edge features RMSE: 0.215 ± 0.055 -> 0.018 ± 0.005 | Improved |
| use JSP-GFN (diag) instead of VBG | joint posterior approximation · d=5 linear Gaussian | edge features Pearson's r: 0.816 ± 0.064 -> 0.998 ± 0.001 | Improved |
| use JSP-GFN (diag) instead of DiBS | parameter posterior · d=5 linear Gaussian | cross-entropy: 9.49 ± 7.34 × 103 -> −4.91 ± 0.51 × 100 | Improved |
| full-covariance Gaussian instead of diagonal in JSP-GFN (ablation) | JSP-GFN · d=5 linear Gaussian | edge features RMSE: 0.018 ± 0.005 -> 0.019 ± 0.007 | No clear effect |
| use JSP-GFN instead of MCMC methods on flow cytometry | JSP-GFN · Flow cytometry | held-out NLL: 2.578 × 104 vs 2.679 × 104 | Mixed |
| use JSP-GFN instead of MH-MC3 on gene expression | JSP-GFN · Gene expression d=61 | held-out NLL: 1.073 × 106 -> 2.651 × 105 | Improved |
| use JSP-GFN instead of Gibbs-MC3 on flow cytometry | JSP-GFN · Flow cytometry | held-out NLL: 3.490 × 105 -> 2.679 × 104 | Improved |
| compare JSP-GFN against baselines on linear Gaussian networks | JSP-GFN · Linear Gaussian d=20 | held-out NLL | Improved |
| compare JSP-GFN against baselines on non-linear Gaussian networks | JSP-GFN · Non-linear Gaussian d=20 | held-out NLL | No clear effect |
| use whole dataset (N=4200) instead of subsample (N=100) for reward (ablation) | JSP-GFN · Flow cytometry | slope of log-prob vs log-reward: 1.0272 -> 0.7817 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DiBS — DIBS: Diversity Inducing Information Bottleneck in Model Ensembles (AAAI 2021) · better on 2 of 2 shared results
- BCD Nets — BCD Nets: Scalable Variational Approaches for Bayesian Causal Discovery (NeurIPS 2021) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On Divergence Measures for Training GFlowNets (NeurIPS 2024 · similarity 0.89)
- Streaming Bayes GFlowNets (NeurIPS 2024 · similarity 0.88)
- Smooth Interpolation for Improved Discrete Graph Generative Models (ICML 2025 · similarity 0.87)
- Revisiting Non-Acyclic GFlowNets in Discrete Environments (ICML 2025 · similarity 0.87)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · similarity 0.87)
- Learning Energy Decompositions for Partial Inference in GFlowNets (ICLR 2024 · similarity 0.87)
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.
