Research map / Probabilistic and scientific modeling / Variational and generative sampling
The Usual Suspects? Reassessing Blame for VAE Posterior Collapse
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 |
|---|---|---|---|
| Increase fully connected network depth (ablation) | AE and VAE · Fashion-MNIST | reconstruction error | Mixed |
| Use a VAE instead of the corresponding deterministic AE | Matching encoder/decoder architectures · Fashion-MNIST and Cifar100 | reconstruction error | Hurt |
| Apply KL warm-start or annealing (ablation) | VAE · Fashion-MNIST and Cifar100 | reconstruction error | Helped |
| Increase convolutional network depth (ablation) | AE and VAE · Cifar100 | reconstruction error | Mixed |
| Add convolution layers inside residual blocks (ablation) | ResNet-inspired AE · Cifar10, Cifar100, SVHN, CelebA | normalized reconstruction error | Mixed |
| Add residual blocks (ablation) | ResNet-inspired AE · Cifar10, Cifar100, SVHN, CelebA | normalized reconstruction error | Hurt |
| Increase convolutional network depth (ablation) | Convolutional AE · ImageNet | reconstruction error | Mixed |
| Increase encoder/decoder depth (ablation) | VAE · CelebA | encoder standard deviation values near 1 | Hurt |
| Introduce a non-affine soft-threshold decoder in place of an affine decoder (ablation) | Gaussian VAE · Constructed data sets | posterior collapse at a bad local minimum | Hurt |
| Allow the decoder noise variance gamma to become sufficiently large | Well-behaved VAE with a non-degenerate decoder · General theoretical setting | posterior collapse | Hurt |
| Learn gamma rather than fixing it to a large value | Affine-decoder VAE · Continuous data | posterior collapse | Helped |
Closest papers by meaning
- Don't Blame the ELBO! A Linear VAE Perspective on Posterior Collapse (NeurIPS 2019 · similarity 0.93)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.93)
- Lagging Inference Networks and Posterior Collapse in Variational Autoencoders (ICLR 2019 · similarity 0.93)
- Marginalization is not Marginal: No Bad VAE Local Minima when Learning Optimal Sparse Representations (ICML 2023 · similarity 0.92)
- Be a Goldfish: Forgetting Bad Conditioning in Sparse Linear Regression via Variational Autoencoders (ICML 2025 · similarity 0.88)
- $t^3$-Variational Autoencoder: Learning Heavy-tailed Data with Student's t and Power Divergence (ICLR 2024 · 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.