Research map / Probabilistic and scientific modeling / Variational and generative sampling
HyperSPNs: Compact and Expressive Probabilistic Circuits
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace weight decay with HyperSPN hypernetwork regularization on Ad (ablation) | RAT-SPN · Twenty Datasets: Ad | test log-likelihood: -54.82 -> -28.58 | Helped |
| Replace weight decay with HyperSPN on BBC (ablation) | RAT-SPN · Twenty Datasets: BBC | test log-likelihood: -262.77 -> -254.44 | Helped |
| Replace weight decay with HyperSPN on Reuters-52 (ablation) | RAT-SPN · Twenty Datasets: Reuters-52 | test log-likelihood: -90.15 -> -86.12 | Helped |
| Replace weight decay with HyperSPN on DNA (ablation) | RAT-SPN · Twenty Datasets: DNA | test log-likelihood: -98.42 -> -98.79 | Hurt |
| Replace weight decay with HyperSPN on MSNBC (ablation) | RAT-SPN · Twenty Datasets: MSNBC | test log-likelihood: -6.04 -> -6.05 | No clear effect |
| Replace weight decay with HyperSPN on Twenty Datasets overall (about 5x fewer parameters) (ablation) | RAT-SPN · Twenty Datasets | test log-likelihood (18/20 datasets) | Helped |
| Replace weight decay with HyperSPN on Amazon Baby Registries (ablation) | RAT-SPN · Amazon Baby Registries | test log-likelihood: -4.54 -> -4.32 | Helped |
| Replace weight decay with HyperSPN on Amazon Gear (ablation) | RAT-SPN · Amazon Baby Registries: Gear | test log-likelihood: -9.21 -> -9.21 | No clear effect |
| Vary weight decay for SPN (1e-3, 1e-4, 1e-5) (ablation) | SPN · Twenty Datasets + Amazon, average | test log-likelihood: -37.18 / -36.91 / -37.17 | Mixed |
| Vary embedding size h (5, 10, 20) (ablation) | HyperSPN · Twenty Datasets + Amazon, average | test log-likelihood: -35.79 / -35.63 / -35.68 | Mixed |
| HyperSPN instead of SPN for sample quality (ablation) | SPN · Parzen window, average | Parzen log-likelihood: -3.185 -> -3.157 | Helped |
| HyperSPN vs SPN-Large and SPN-Small on hand-crafted dataset (ablation) | SPN · Hand-crafted toy dataset | test log-likelihood: -166.90 -> -166.32 | Helped |
Closest papers by meaning
- Scaling Probabilistic Circuits via Monarch Matrices (ICML 2025 · similarity 0.87)
- Sum of Squares Circuits (AAAI 2025 · similarity 0.87)
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.86)
- Probabilistic Neural Circuits (AAAI 2024 · similarity 0.86)
- Scaling Tractable Probabilistic Circuits: A Systems Perspective (ICML 2024 · similarity 0.86)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · 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.