Research map / Probabilistic and scientific modeling / Variational and generative sampling
JEM++: Improved Techniques for Training JEM
Xiulong Yang, Shihao Ji
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 |
|---|---|---|---|
| add proximity constraint (Proximal SGLD) (ablation) | JEM++ · CIFAR10 ablation (M=10, N=5) | accuracy (%): 92.9 -> 93.5 | Improved |
| add proximity constraint (ablation) | JEM++ · CIFAR10 ablation (M=10, N=5) | FID: 36.0 -> 37.1 | Hurt |
| use SGD with large learning rate instead of Adam (ablation) | JEM++ · CIFAR10 ablation (M=10, N=5) | FID: 42.7 -> 37.1 | Improved |
| use SGD instead of Adam (ablation) | JEM++ · CIFAR10 ablation (M=10, N=5) | accuracy (%): 92.5 -> 93.5 | Improved |
| add informative initialization with batch norm (ablation) | JEM++ · CIFAR10 ablation (M=10, N=5) | accuracy (%): 88.6 -> 93.5 | Improved |
| enable batch norm with informative init (ablation) | JEM++ · CIFAR10 ablation (M=10, N=5) | accuracy (%): 91.1 -> 93.5 | Improved |
| increase SGLD steps M from 5 to 10 | JEM++ · CIFAR10 | accuracy (%): 91.1 -> 93.5 | Improved |
| increase M from 10 to 20 | JEM++ · CIFAR10 | accuracy (%): 93.5 -> 94.1 | Improved |
| increase M from 10 to 20 | JEM++ · CIFAR10 | per-epoch speedup vs JEM: 1.49x -> .93x | Hurt |
| increase inner steps N from 5 to 20 | JEM++ (M=10) · CIFAR10 | minutes per epoch: 20.1 -> 20.4 | No clear effect |
| replace JEM with JEM++ (M=10) | JEM (K=20) · CIFAR100 | accuracy (%): 72.2 -> 74.5 | Improved |
| replace JEM with JEM++ (M=10) | JEM (K=20) · SVHN | accuracy (%): 96.7 -> 96.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Glow — Glow: Generative Flow with Invertible 1x1 Convolutions (NeurIPS 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Bridging the Performance Gaps of Joint Energy-Based Models (CVPR 2023 · similarity 0.90)
- Your Classifier Can Do More: Towards Balancing the Gaps in Classification, Robustness, and Generation (CVPR 2026 · similarity 0.89)
- Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation (ICLR 2026 · similarity 0.89)
- Your classifier is secretly an energy based model and you should treat it like one (ICLR 2020 · similarity 0.84)
- MARS: Unleashing the Power of Variance Reduction for Training Large Models (ICML 2025 · similarity 0.84)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.83)
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.
