Research map / Probabilistic and scientific modeling / Variational and generative sampling
Structure by Architecture: Structured Representations without Regularization
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 |
|---|---|---|---|
| use a structural decoder that injects each latent variable at a separate layer without latent regularization | Autoencoder · 3D-Shapes, MPI3D-Toy, MPI3D-Sim, MPI3D-Real | DCI | Helped |
| use 12 structural variables rather than 6 (ablation) | SAE · 3D-Shapes | DCI: 0.865 -> 0.974 | Helped |
| use 12 structural variables rather than 6 (ablation) | SAE · MPI3D-Toy | DCI: 0.454 -> 0.642 | Helped |
| use 12 structural variables rather than 6 (ablation) | SAE · MPI3D-Sim | DCI: 0.294 -> 0.411 | Helped |
| use 12 structural variables rather than 6 (ablation) | SAE · MPI3D-Real | DCI: 0.295 -> 0.374 | Helped |
| give each decoder layer access to all latent variables rather than only its designated variable (ablation) | SAE decoder · 3D-Shapes | DCI: 0.974 -> 0.299 | Hurt |
| use hybrid sampling instead of sampling from the matched prior | VAE-based models · Image generation across evaluated datasets | FID | Helped |
| use the structural decoder architecture rather than an unstructured hourglass architecture | Autoencoder · Evaluated datasets | reconstruction quality | Helped |
| update the decoder rather than only the encoder on novel observations | Autoencoder · 3D-Shapes extrapolation | reconstruction quality | Helped |
| use AdaAE rather than SAE (ablation) | Autoencoder · Celeb-A hybrid sampling | FID | Helped |
Closest papers by meaning
- An Architecture for Deep, Hierarchical Generative Models (NeurIPS 2016 · similarity 0.88)
- Unbiased learning of deep generative models with structured discrete representations (NeurIPS 2023 · similarity 0.87)
- Decoupling Global and Local Representations via Invertible Generative Flows (ICLR 2021 · similarity 0.87)
- Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.86)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.86)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.86)
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.