Research map / Probabilistic and scientific modeling / Variational and generative sampling
VAEBM: A Symbiosis between Variational Autoencoders and Energy-based Models
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 a pixel-space energy model to the VAE generator | NVAE on CIFAR-10 · CIFAR-10 | FID: 51.67 -> 12.26 | Helped |
| Use persistent chains during VAEBM training (ablation) | VAEBM · CIFAR-10 | FID: 12.26 -> 12.19 | Helped |
| Initialize data-space EBM MCMC with NVAE samples instead of noise (ablation) | EBM on x · CIFAR-10 | FID: 48.89 -> 29.32 | Helped |
| Define the EBM over augmented data-and-latent space rather than data space only (ablation) | EBM initialized with NVAE samples · CIFAR-10 | FID: 29.32 -> 12.96 | Helped |
| Train the VAE decoder adversarially using the WGAN objective (ablation) | NVAE decoder with energy network · CIFAR-10 | FID: 12.96 vs 20.39 | Hurt |
| Update the VAE decoder during the EBM stage with the additional KL objective (ablation) | VAEBM · CIFAR-10 | FID: 12.96 -> 14.0 | No clear effect |
| Add the EBM refinement component to the VAE | VAE on 25-Gaussians · 25-Gaussians test | average log likelihood (nats): -2.97 -> -1.50 | Helped |
| Add the energy function to the VAE model | NVAE trained on CIFAR-10 · CIFAR-10 OOD: SVHN | AUROC: 0.42 -> 0.83 | Helped |
| Reparameterize Langevin updates in the VAE noise variables instead of using a universal step size in data-and-latent space (ablation) | VAEBM sampling · Image generation | sample quality | Helped |
| Use persistent sampling during training (ablation) | VAEBM on CelebA HQ 256 · CelebA HQ 256 | FID | Mixed |
Closest papers by meaning
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.90)
- On the Value of Infinite Gradients in Variational Autoencoder Models (NeurIPS 2021 · similarity 0.89)
- Marginalization is not Marginal: No Bad VAE Local Minima when Learning Optimal Sparse Representations (ICML 2023 · similarity 0.88)
- NVAE: A Deep Hierarchical Variational Autoencoder (NeurIPS 2020 · similarity 0.88)
- Embrace the Gap: VAEs Perform Independent Mechanism Analysis (NeurIPS 2022 · similarity 0.88)
- A Contrastive Learning Approach for Training Variational Autoencoder Priors (NeurIPS 2021 · similarity 0.88)
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.