Research map / Probabilistic and scientific modeling / Variational and generative sampling
Conjugate Energy-Based Models
Hao Wu, Babak Esmaeili, Michael Wick, Jean-Baptiste Tristan, Jan-Willem Van De Meent
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 |
|---|---|---|---|
| Replace the VAE-style pixel reconstruction objective with a conjugate energy over learned latent sufficient statistics | CEBM versus VAE · CIFAR-10 test set, full training set | classification accuracy: 38 -> 53 | Improved |
| Replace the VAE-style pixel reconstruction objective with a conjugate energy over learned latent sufficient statistics | CEBM versus VAE · SVHN test set, full training set | classification accuracy: 36 -> 70 | Improved |
| Add a mixture-of-Gaussians bias instead of a spherical Gaussian bias (ablation) | CEBM · MNIST test set, full training set | classification accuracy: 97 -> 98 | Improved |
| Add a mixture-of-Gaussians bias instead of a spherical Gaussian bias (ablation) | CEBM · Fashion-MNIST test set, full training set | classification accuracy: 83 -> 85 | Improved |
| Add a mixture-of-Gaussians bias instead of a spherical Gaussian bias (ablation) | CEBM · CIFAR-10 test set, 1 example per class | classification accuracy: 19 -> 16 | Hurt |
| Add a mixture-of-Gaussians bias instead of a spherical Gaussian bias (ablation) | CEBM · SVHN test set, full training set | classification accuracy: 70 -> 60 | Hurt |
| Add a mixture-of-Gaussians bias instead of a spherical Gaussian bias (ablation) | CEBM · Fashion-MNIST-trained; MNIST OOD; log density | AUROC: .37 -> .56 | Improved |
| Add a mixture-of-Gaussians bias instead of a spherical Gaussian bias (ablation) | CEBM · CIFAR-10-trained; SVHN OOD; log density | AUROC: .47 -> .55 | Improved |
| Use the CEBM learned representation rather than pixel distance for nearest-neighbor similarity | CEBM · CIFAR-10 nearest-neighbor class agreement | same-class neighbor fraction: 0.38 -> 0.45 | Improved |
| Use a conjugate EBM representation rather than a VAE representation | CEBM versus VAE · MNIST aggregate posterior | mutual information: 9.1 -> 0.3 | Hurt |
Closest papers by meaning
- Guiding Energy-based Models via Contrastive Latent Variables (ICLR 2023 · similarity 0.88)
- Learning Latent Space Energy-Based Prior Model (NeurIPS 2020 · similarity 0.88)
- VAEBM: A Symbiosis between Variational Autoencoders and Energy-based Models (ICLR 2021 · similarity 0.88)
- Energy-Based Modelling for Discrete and Mixed Data via Heat Equations on Structured Spaces (NeurIPS 2024 · similarity 0.87)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.87)
- Learning Latent Space Hierarchical EBM Diffusion Models (ICML 2024 · similarity 0.85)
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.
