Research map / Image generation / Autoregressive and latent generation
Neurosymbolic Deep Generative Models for Sequence Data with Relational Constraints
Halley Young, Maxwell Du, Osbert Bastani
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 |
|---|---|---|---|
| remove program synthesis (use full adjacency tensor) (ablation) | SGM A2 · Essen music | Frechet distance: 32.7 vs 42.1 (without synthesis) | Improved |
| remove program synthesis (ablation) | SGM A2 · Essen music | GCN discriminator loss: 0.63 vs 0.42 (without synthesis) | Improved |
| replace constraint-solving synthesis with greedy synthesis (ablation) | SGM A2 · Essen music | Frechet distance: 32.7 vs 40.5 (greedy) | Improved |
| replace greedy synthesis with constraint solving (ablation) | SGM A2 · Essen music | GCN discriminator loss: 0.50 -> 0.63 | Improved |
| use continuous cosine-similarity relation instead of discrete relations (ablation) | SGM A2 · Essen music | Frechet distance: 32.7 vs 33.2 | No clear effect |
| use continuous relation (ablation) | SGM A2 · Essen music | GCN discriminator loss: 0.63 -> 0.46 | Hurt |
| sampling strategy A2 vs MusicVAE baseline | SGM · Essen music | NLL: 1158.6 -> 1028.4 | Improved |
| sampling strategy A2 vs MusicAutobot baseline | SGM · Essen music | Frechet distance: 53.7 -> 32.7 | Improved |
| use A1 instead of A2 | SGM · Essen music | Frechet distance: 32.7 -> 43.4 | Hurt |
| use A3 instead of A2 | SGM · Essen music | GCN discriminator loss: 0.63 -> 0.43 | Hurt |
| sample constraints uniformly instead of learned distribution (ablation) | SGM A1 · Poetry | GCN discriminator loss: 0.59 -> 0.69 | Improved |
| use A3 (combinatorial optimization) instead of A1 | SGM · Poetry | Frechet distance: 21.5 -> 13.51 | Improved |
Closest papers by meaning
- How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.86)
- Relational Neurosymbolic Markov Models (AAAI 2025 · similarity 0.85)
- Towards a theory of how the structure of language is acquired by deep neural networks (NeurIPS 2024 · similarity 0.84)
- Whole-Song Hierarchical Generation of Symbolic Music Using Cascaded Diffusion Models (ICLR 2024 · similarity 0.84)
- Controllable Generation via Locally Constrained Resampling (ICLR 2025 · similarity 0.84)
- Neurosymbolic Grounding for Compositional World Models (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.
