Research map / Probabilistic and scientific modeling / Variational and generative sampling
On the Latent Variable Interpretation in Sum-Product Networks
Robert Peharz, Robert Gens, Franz Pernkopf, Pedro Domingos
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 |
|---|---|---|---|
| update Gaussian variances in EM (V, WV, MV, WMV) (ablation) | EM post-training of original SPN parameters · Caltech-101+ORL (103 datasets, test, EM post-training of original params) | test log-likelihood: -13.47 vs 0.55 (after 1st iteration) | Hurt |
| update sum-weights and means (WM) vs only weights (W) (ablation) | EM post-training · Caltech-101+ORL (103 datasets, test, EM post-training of original params) | relative test log-likelihood after 1st iteration: 0.55 -> 1.06 | Improved |
| post-train original hard-EM parameters with standard EM (W, M, WM) | SPN trained by hard EM · Caltech-101+ORL (103 datasets, test, EM post-training of original params) | test log-likelihood: 0.58% (M) up to 1.39% (WM) | Improved |
| use soft EM with random init of sum-weights on oracle structure | SPN with PD structure · Caltech-101+ORL (103 datasets, W trained from random init) train | log-likelihood vs original parameters: 100.00% of datasets higher, 3.97% gain | Improved |
| use soft EM with random init of sum-weights on oracle structure | SPN with PD structure · Caltech-101+ORL (103 datasets, W trained from random init) test | log-likelihood vs original parameters: 67.96% of datasets higher at best iteration; 41.75% after 1st iteration | Mixed |
| use uniform instead of deterministic twin-weights in MPE | MPESelective on augmented SPN, 16 RVs, alpha=1.0 · augmented SPN with uniform twin-weights | MPE log-likelihood gap: -0.85 (12) -> 0.00 (100) | Improved |
| use deterministic twin-weights in MPE | MPESelective on augmented SPN, 16 RVs, alpha=1.0 · augmented SPN with deterministic twin-weights | MPE log-likelihood gap: -1.11 (12) -> 0.00 (100) | Improved |
| MPEDet vs MPEUni as approximation for MPE in original SPNs | MPESelective · original SPN, 16 RVs, alpha=1.0 | MPE log-likelihood gap: -0.76 (3) vs -1.18 (2) | No clear effect |
Closest papers by meaning
- Bayesian Learning of Sum-Product Networks (NeurIPS 2019 · similarity 0.91)
- Sum-product networks: A survey (TPAMI 2021 · similarity 0.90)
- Maximum A Posteriori Inference in Sum-Product Networks (AAAI 2018 · similarity 0.89)
- Inferring Dynamic Networks from Marginals with Iterative Proportional Fitting (ICML 2024 · similarity 0.84)
- Sum-Product-Set Networks: Deep Tractable Models for Tree-Structured Graphs (ICLR 2024 · similarity 0.84)
- Tractable Probabilistic Graph Representation Learning with Graph-Induced Sum-Product Networks (ICLR 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.
