Research map / Probabilistic and scientific modeling / Variational and generative sampling
Guiding Energy-based Models via Contrastive Latent Variables
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 CLEL (contrastive latent variable guidance) to base EBM | EBM on CIFAR-10 · CIFAR-10 unconditional | FID: 23.50 -> 15.27 | Helped |
| add latent variable p(z|x) learning with MLP projector (vs beta=0 baseline) (ablation) | small ResNet latent-variable EBM · CIFAR-10 component ablation | FID: 42.46 -> 36.29 | Helped |
| use generated latent samples as additional negatives in contrastive loss (ablation) | small ResNet latent-variable EBM · CIFAR-10 component ablation | FID: 36.29 -> 35.73; OOD 0.8580 -> 0.8723 | Helped |
| remove MLP projector (identity projection) (ablation) | latent-variable EBM · CIFAR-10 component ablation | FID: 35.73 -> 86.02 | Hurt |
| 2-layer MLP projector replaced by linear projector (ablation) | latent-variable EBM · CIFAR-10 component ablation | FID: 35.73 -> 37.35 | Hurt |
| increase latent coupling beta from 0.01 to 0.1 (ablation) | latent-variable EBM · CIFAR-10 beta sensitivity | FID: 35.73 -> 56.39 | Hurt |
| replace SimCLR with BYOL as latent encoder (ablation) | CLEL latent encoder · CIFAR-10 component ablation | FID: 35.73 -> 36.31 (OOD 0.8723 -> 0.8792 slightly higher) | Hurt |
| replace SimCLR with MAE as latent encoder (ablation) | CLEL latent encoder · CIFAR-10 component ablation | FID: 35.73 -> 37.67 | Hurt |
| add multi-scale architecture (base to larger multi-scale EBM) | CLEL on CIFAR-10 · CIFAR-10 unconditional | FID: 15.27 -> 12.46 | Helped |
| CRL batch size 256 for encoder (vs base) | CLEL on CIFAR-10 · CIFAR-10 unconditional | FID: 12.46 -> 11.65 | Helped |
| CLEL-Large with more channels | CLEL on CIFAR-10 · CIFAR-10 unconditional | FID: 11.65 -> 8.61 | Helped |
| CLEL with OOD score using joint density (adds latent term) | CLEL-Base OOD detection · CIFAR-10 OOD SVHN AUROC | AUROC: 0.9848 vs Improved CD 0.91 | Helped |
Closest papers by meaning
- Energy-Based Contrastive Learning of Visual Representations (NeurIPS 2022 · similarity 0.89)
- Latent-Guided Cooperative Energy-Based Models (ICML 2026 · similarity 0.89)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.89)
- Improved Contrastive Divergence Training of Energy-Based Models (ICML 2021 · similarity 0.89)
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.88)
- Learning Latent Space Hierarchical EBM Diffusion Models (ICML 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.