Research map / Probabilistic and scientific modeling / Variational and generative sampling
Disentangling Disentanglement in 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 |
|---|---|---|---|
| Replace the isotropic Gaussian prior with a non-isotropic Gaussian prior | β-VAE on 2D Shapes · 2D Shapes | disentanglement metric | Helped |
| Learn the non-isotropic prior variance | β-VAE with non-isotropic Gaussian prior · 2D Shapes | disentanglement metric | Helped |
| Replace the isotropic Gaussian prior with a product Student-t prior | β-VAE · 2D Shapes | disentanglement metric | Helped |
| Reduce the Student-t degrees of freedom to very low values (ablation) | β-VAE with product Student-t prior · 2D Shapes | disentanglement metric | Mixed |
| Increase β while retaining an isotropic Gaussian prior (ablation) | β-VAE · 2D Shapes | disentanglement metric | No clear effect |
| Increase β (ablation) | Lα,β with four-component Gaussian mixture prior · Pinwheel | latent overlap | Helped |
| Increase β too far (ablation) | Lα,β with four-component Gaussian mixture prior · Pinwheel | meaningfulness of datapoint encodings | Hurt |
| Increase α (ablation) | Lα,β with four-component Gaussian mixture prior · Pinwheel | aggregate posterior-prior match | Helped |
| Replace a Gaussian prior (γ=0) with the sparse prior (γ=0.8) (ablation) | Lα,β on Fashion-MNIST · Fashion-MNIST | Hoyer sparsity | Helped |
| Add aggregate-posterior regularization (α>0) (ablation) | Lα,β with sparse prior · Fashion-MNIST | Hoyer sparsity | Helped |
| Increase α to large values (ablation) | Lα,β with sparse prior · Fashion-MNIST | reconstruction | Hurt |
| Set β too low or too high (ablation) | Lα,β with sparse prior · Fashion-MNIST | Hoyer sparsity | Hurt |
Closest papers by meaning
- Isolating Sources of Disentanglement in Variational Autoencoders (NeurIPS 2018 · similarity 0.94)
- Orthogonality-Enforced Latent Space in Autoencoders: An Approach to Learning Disentangled Representations (ICML 2023 · similarity 0.93)
- Local Disentanglement in Variational Auto-Encoders Using Jacobian $L_1$ Regularization (NeurIPS 2021 · similarity 0.93)
- An Identifiable Double VAE For Disentangled Representations (ICML 2021 · similarity 0.92)
- $\alpha$TC-VAE: On the relationship between Disentanglement and Diversity (ICLR 2024 · similarity 0.90)
- Tripod: Three Complementary Inductive Biases for Disentangled Representation Learning (ICML 2024 · similarity 0.90)
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.