Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learning Manifold Dimensions with Conditional Variational Autoencoders
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 kappa=20 (>= r) instead of kappa=5 (< r) latent dimensions (ablation) | VAE · Synthetic d=20, r=10 | -ELBO: -16.96 -> -28.34 | Helped |
| condition on label (CVAE) instead of unconditional VAE | MNIST kappa=32 · MNIST | -ELBO: -1224.37 -> -1489.42 | Helped |
| condition on label (CVAE) instead of unconditional VAE | Fashion MNIST kappa=32 · Fashion MNIST | -ELBO: -1327.26 -> -1239.09 | No clear effect |
| increase effective dimension t of the conditioning variable (ablation) | CVAE · Synthetic r=10 | active dimensions: 9 -> 1 (t=1 to t=9) | Helped |
| add attention layer to CVAE decoder (ablation) | CVAE · Union of 5 manifolds (discrete c) | active dimensions: 4 -> 1 (r=1); 4 -> 5 (r=5) | Helped |
| add attention layer to CVAE decoder (ablation) | CVAE · Union of 5 manifolds (discrete c) | -ELBO: -7.58 -> -59.25 (r=5) | Helped |
| add attention layer to CVAE decoder (ablation) | CVAE · Continuous c, r=12 | -ELBO: -9.69 -> -41.49 (t=2); -61.77 -> -55.14 (t=10) | Mixed |
| initialize decoder variance log gamma larger (20 vs -20) (ablation) | VAE · Synthetic d=20, r=10 | active dimensions: 10 -> 1 | Hurt |
| initialize decoder variance log gamma larger (20 vs -20) (ablation) | CVAE · Synthetic d=20, r=10, t=5 | active dimensions: 5 -> 5 | No clear effect |
| use parameterized conditional prior instead of N(0,I) prior (ablation) | CVAE · Synthetic d=20, r=10, t=5 | -ELBO: -44.38 -> -45.2 (log gamma 0) | No clear effect |
| share weights between encoder and prior (ablation) | sequential CVAE · ARMA sequences | -ELBO: -45.015 -> -2.49 | Hurt |
| use full covariance encoder instead of diagonal (ablation) | VAE · MNIST kappa=32 | -ELBO: -1205 -> -1756 | Helped |
Closest papers by meaning
- Variational autoencoders in the presence of low-dimensional data: landscape and implicit bias (ICLR 2022 · similarity 0.91)
- Mixed-curvature Variational Autoencoders (ICLR 2020 · similarity 0.87)
- Consistency Regularization for Variational Auto-Encoders (NeurIPS 2021 · similarity 0.87)
- A Geometric Perspective on Variational Autoencoders (NeurIPS 2022 · similarity 0.87)
- Variational Autoencoded Regression: High Dimensional Regression of Visual Data on Complex Manifold (CVPR 2017 · similarity 0.87)
- Sparse Autoencoders, Again? (ICML 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.