Research map / Probabilistic and scientific modeling / Variational and generative sampling
Be More Active! Understanding the Differences Between Mean and Sampled Representations of Variational Autoencoders
Lisa Bonheme, Marek Grzes
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 passive variables from the mean representation before computing total correlation (ablation) | Mean representations of VAEs (beta-TC VAE) · noisy dSprites | total correlation | Improved |
| remove passive variables from the mean representation before computing averaged mutual information (ablation) | Mean representations of VAEs (DIP-VAE II) · dSprites | averaged mutual information | Improved |
| remove mixed variables in addition to passive ones from the mean representation (ablation) | Mean representations of VAEs · dSprites and noisy dSprites | TC and averaged MI | No clear effect |
| add passive variables to the sampled representation | Sampled representations of VAEs · dSprites, smallNorb, Cars3D, colour dSprites | total correlation | No clear effect |
| add passive variables to the sampled representation | Sampled representations of VAEs · Cars3D | averaged mutual information | Hurt |
| increase regularisation strength (beta) | VAE latent representations · disentanglement_lib models, 6 datasets | number of passive and mixed variables | Improved |
| use only passive variables of the sampled representation for logistic regression (ablation) | beta-VAE · dSprites | logistic regression accuracy | Hurt |
| use only passive variables of the mean representation with increasing regularisation (ablation) | beta-VAE · dSprites | logistic regression accuracy | Improved |
| raise channel capacity C in annealed VAE | Annealed VAE mean representation · Cars3D | effective rank vs passive variable count: channel capacity of 25 and 4 passive variables is the same as 100 and 3 passive variables | Mixed |
Closest papers by meaning
- Disentangling Disentanglement in Variational Autoencoders (ICML 2019 · similarity 0.90)
- Isolating Sources of Disentanglement in Variational Autoencoders (NeurIPS 2018 · similarity 0.89)
- Exploring the Latent Space of Autoencoders with Interventional Assays (NeurIPS 2022 · similarity 0.88)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.88)
- $\alpha$TC-VAE: On the relationship between Disentanglement and Diversity (ICLR 2024 · similarity 0.87)
- Synergy Between Sufficient Changes and Sparse Mixing Procedure for Disentangled Representation Learning (ICLR 2025 · similarity 0.87)
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.
