Research map / Probabilistic and scientific modeling / Variational and generative sampling
Sum-Product Autoencoding: Encoding and Decoding Representations Using Sum-Product Networks
Antonio Vergari, Robert Peharz, Nicola Di Mauro, Alejandro Molina, Kristian Kersting, Floriana Esposito
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 |
|---|---|---|---|
| encode label space with SPAE CAT embeddings (scenario II) rather than raw labels | Linear predictor on MLC · 10 MLC benchmarks | avg relative JAC vs LR: +24.07 | Improved |
| encode features with SPAE CAT embeddings (scenario I) | LR · 10 MLC benchmarks | avg relative JAC vs LR: -11.90 | Hurt |
| encode features with SPAE ACT embeddings (scenario I) | LR · 10 MLC benchmarks | avg relative EXA vs LR: +17.18 | Improved |
| combine ACT feature embedding with CAT label embedding (scenario III) (ablation) | SPAE · 10 MLC benchmarks | avg relative JAC vs LR: +24.07 -> +25.47 | Improved |
| use kNN decoding instead of tree decoding gCAT (ablation) | SPAE CAT · MLC scenario II | avg relative JAC vs LR: +24.07 -> +22.83 | Hurt |
| use kNN decoding for MADE decoder (ablation) | MADE · MLC scenario II | avg relative JAC vs LR: -30.42 -> +14.57 | Improved |
| remove up to half of label embedding components at random (ablation) | SPAE CAT/ACT decoding · EXA | EXA: degrades by less than 50% on average when half missing | Hurt |
Closest papers by meaning
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.85)
- Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression (NeurIPS 2026 · similarity 0.85)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.85)
- Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders (ICML 2025 · similarity 0.85)
- Tractable Probabilistic Graph Representation Learning with Graph-Induced Sum-Product Networks (ICLR 2024 · similarity 0.85)
- Sparse Autoencoders, Again? (ICML 2025 · 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.
