Research map / Probabilistic and scientific modeling / Variational and generative sampling
Preventing Posterior Collapse with delta-VAEs
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 |
|---|---|---|---|
| Add an auxiliary prior (ablation) | δ-VAE + PixelSNAIL · CIFAR-10 test | negative log-likelihood (bits/dim): ≤ 2.85 (0.02) -> ≤ 2.83 (0.01) | Helped |
| Add an auxiliary prior (ablation) | δ-VAE + PixelSNAIL · ImageNet 32×32 valid | negative log-likelihood (bits/dim): ≤ 3.78 (0.08) -> ≤ 3.77 (0.07) | Helped |
| Add an auxiliary prior | δ-VAE with Transformer decoder · LM1B | negative log-likelihood (nats/token): ≤ 3.61 (0.2) -> ≤ 3.58(0.17) | Helped |
| Use a temporal AR(1) prior instead of an independent Gaussian posterior setup (ablation) | δ-VAE (δ=0.08) · CIFAR-10 test | negative log-likelihood (table's Test ELBO): 3.08 (0.08) -> 3.02 (0.09) | Helped |
| Use a temporal AR(1) prior instead of an independent Gaussian posterior setup (ablation) | δ-VAE (δ=0.08) · CIFAR-10 | linear classification accuracy: 66% -> 65% | Hurt |
| Use an anti-causal encoder instead of a non-causal encoder (ablation) | Temporal δ-VAE with AR(1) prior · CIFAR-10, low-rate, 6-layer decoder with no attention | negative log-likelihood: 3.04 (0.02) -> 3.07 (0.02) | Hurt |
| Use an anti-causal encoder instead of a non-causal encoder (ablation) | Temporal δ-VAE with AR(1) prior · CIFAR-10, low-rate, 8-layer decoder with attention | negative log-likelihood: 3.01 (0.03) -> 3.01 (0.03) | No clear effect |
| Use an anti-causal encoder instead of a non-causal encoder (ablation) | Temporal δ-VAE with AR(1) prior · CIFAR-10, high-rate, 8-layer decoder with attention | negative log-likelihood: 3.32 (0.22) -> 3.22 (0.22) | Helped |
| Use an anti-causal encoder instead of a non-causal encoder (ablation) | Temporal δ-VAE with AR(1) prior · CIFAR-10, low-rate, large decoder | negative log-likelihood: 2.88 (0.05) -> 2.87 (0.05) | Helped |
| Use an anti-causal encoder instead of a non-causal encoder (ablation) | Temporal δ-VAE with auxiliary prior · CIFAR-10, high-rate, 8-layer decoder with attention | negative log-likelihood: 3.11 (0.02) -> 3.03 (0.03) | Helped |
| Use an anti-causal encoder instead of a non-causal encoder (ablation) | Temporal δ-VAE with auxiliary prior · CIFAR-10, low-rate, 6-layer decoder with no attention | negative log-likelihood: 3.03 (0.01) -> 3.06 (0.01) | Hurt |
| Use anti-causal rather than non-causal encoder structure (ablation) | Temporal δ-VAE · CIFAR-10 | posterior collapse: < 10^{-8} bits per dimension | No clear effect |
Closest papers by meaning
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.92)
- Lagging Inference Networks and Posterior Collapse in Variational Autoencoders (ICLR 2019 · similarity 0.92)
- Posterior Collapse and Latent Variable Non-identifiability (NeurIPS 2021 · similarity 0.91)
- The Usual Suspects? Reassessing Blame for VAE Posterior Collapse (ICML 2020 · similarity 0.90)
- Don't Blame the ELBO! A Linear VAE Perspective on Posterior Collapse (NeurIPS 2019 · similarity 0.90)
- EQ-VAE: Equivariance Regularized Latent Space for Improved Generative Image Modeling (ICML 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.