Research map / Learning theory and algorithms / Kernels and regression theory
Exploring Bidirectional Bounds for Minimax-Training of Energy-Based Models
Cong Geng, Jia Wang, Li Chen, Zhiyong Gao, Jes Frellsen, Søren Hauberg
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 zero-centered gradient penalty with the derived GP upper bound (ablation) | EBM SV lower bound · ANIMEFACE | FID: 20.53 -> 12.75 | Improved |
| replace zero-centered gradient penalty with the derived GP upper bound (ablation) | EBM SV lower bound · CIFAR-10 | FID: 35.42 -> 28.63 | Improved |
| use a true upper bound instead of zero-centered GP (ablation) | EBM generator · ANIMEFACE | anisotropy index: 3.4560 -> 1.8016 | Improved |
| use GP upper bound instead of zero-centered GP with SV lower bound (ablation) | EBM on normalizing flow · MNIST | test log-likelihood: -2167 -> -1634 | Improved |
| use GP upper bound instead of zero-centered GP with MI lower bound (ablation) | EBM on normalizing flow · MNIST | test log-likelihood: -1819 -> -1782 | Improved |
| use diffusion upper bound with SV lower bound instead of GP (ablation) | EBM on normalizing flow · MNIST | test log-likelihood: -1634 -> -1561 | Improved |
| swap MI lower bound for SV lower bound (GP upper) (ablation) | EBM · CIFAR-10 | FID: 28.96 -> 28.63 | No clear effect |
| raise entropy weight lambda from 0.0001 to 1 (ablation) | EBM SV+GP · CIFAR-10 | FID: 28.63 -> 45.08 | Hurt |
| raise entropy weight lambda from 0.0001 to 1 (ablation) | EBM SV+Diff · CIFAR-10 | FID: 29.31 -> 33.13 | Hurt |
| scale network from DCGAN to UNet | EBM MI+Diff · CIFAR-10 | FID: 30.37 -> 5.47 | Improved |
| scale to large UNet | EBM MI+Diff · ImageNet 32x32 | FID: 14.61 -> 6.57 | Improved |
| add bidirectional bounds (SV+GP) versus VERA lower-bound-only minimax | EBM · CIFAR-10 | FID: 66.38 -> 28.63 | Improved |
Closest papers by meaning
- Bounds all around: training energy-based models with bidirectional bounds (NeurIPS 2021 · similarity 0.97)
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.89)
- Learning Energy-Based Models with Adversarial Training (ECCV 2022 · similarity 0.88)
- Energy-Inspired Models: Learning with Sampler-Induced Distributions (NeurIPS 2019 · similarity 0.87)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.85)
- Joint Learning of Energy-based Models and their Partition Function (ICML 2025 · similarity 0.84)
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.
