Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learning Joint Latent Space EBM Prior Model for Multi-Layer Generator
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 |
|---|---|---|---|
| Use a joint latent-space EBM prior instead of the Gaussian prior | Multi-layer generator on SVHN, CelebA-64, and CIFAR-10 · SVHN, CelebA-64, CIFAR-10 | FID: 42.02 -> 26.81; 40.95 -> 33.60; 74.16 -> 66.32 | Helped |
| Add the joint EBM prior to the NVAE backbone | NVAE · CIFAR-10 | FID: 37.73 -> 11.34 | Helped |
| Add the joint EBM prior to the NVAE backbone | NVAE · CelebA-HQ-256 and LSUN-Church-64 | FID: 30.25 -> 9.89; 38.13 -> 8.38 | Helped |
| Use a joint EBM prior for anomaly detection instead of a single-layer latent EBM prior | MNIST anomaly detection · Held-out digits 1, 4, 5, 7, and 9 | AUPRC: 0.336 ± 0.008 -> 0.470 ± 0.009; 0.630 ± 0.017 -> 0.941 ± 0.001; 0.619 ± 0.013 -> 0.964 ± 0.003; 0.463 ± 0.009 -> 0.815 ± | Helped |
| Replace the joint EBM prior with a Gaussian prior, even while increasing generator parameters (ablation) | Multi-layer generator · Gaussian-prior comparison setting | FID: 28.60 -> 42.03; 28.60 -> 39.82; 28.60 -> 37.75; 28.60 -> 36.10 | Hurt |
| Increase EBM hidden units from nef = 10 to nef = 100 (ablation) | Joint EBM prior · CIFAR-10 | FID: 69.73 -> 66.32 | Helped |
| Replace posterior MCMC sampling with an inference model (ablation) | Joint EBM prior · CIFAR-10 | FID: 66.32 -> 68.45 | Hurt |
| Replace posterior MCMC sampling with an inference model (ablation) | Joint EBM prior · CIFAR-10 | Time(s) per iteration: 0.568 -> 0.256 | Helped |
| Increase prior Langevin sampling steps from 10 to 40 (ablation) | Joint EBM prior · CIFAR-10 | FID: 69.42 -> 66.32 | Helped |
| Increase prior Langevin sampling steps from 40 to 100 (ablation) | Joint EBM prior · CIFAR-10 | FID: 66.32 -> 65.86 | No clear effect |
| Use the joint EBM prior rather than the BIVA backbone decision function | BIVA OOD detector · CIFAR-10 in-distribution, SVHN out-of-distribution | AUROC: 0.885 -> 0.927 | Helped |
| Use the likelihood-ratio decision function rather than the log-posterior at k = 9 (ablation) | BIVA and joint EBM prior OOD detection · CIFAR-10 in-distribution, SVHN out-of-distribution | AUROC: 0.870 -> 0.927 | Helped |
Closest papers by meaning
- Learning Latent Space Energy-Based Prior Model (NeurIPS 2020 · similarity 0.93)
- Learning Latent Space Hierarchical EBM Diffusion Models (ICML 2024 · similarity 0.92)
- Learning Probabilistic Models from Generator Latent Spaces with Hat EBM (NeurIPS 2022 · similarity 0.89)
- Joint Training of Variational Auto-Encoder and Latent Energy-Based Model (CVPR 2020 · similarity 0.88)
- Adaptive Multi-stage Density Ratio Estimation for Learning Latent Space Energy-based Model (NeurIPS 2022 · similarity 0.88)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.86)
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.