Research map / Probabilistic and scientific modeling / Variational and generative sampling
Pseudo-Spherical Contrastive Divergence
Lantao Yu, Jiaming Song, Yang Song, Stefano Ermon
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 |
|---|---|---|---|
| Use PS-CD with γ=1.0 instead of Contrastive Divergence | EBM image generation · CIFAR-10 (32 × 32) conditional | FID: 37.90 -> 29.78 | Improved |
| Use PS-CD with γ=-0.5 instead of Contrastive Divergence | EBM image generation · CelebA (64 × 64) | FID: 26.10 -> 20.35 | Improved |
| Set γ to 1.0 rather than 2.0 (ablation) | PS-CD · CIFAR-10 (32 × 32) conditional | FID: 33.19 -> 29.78 | Improved |
| Set γ to -0.5 rather than 1.0 (ablation) | PS-CD · CIFAR-10 (32 × 32) conditional | FID: 29.78 -> 27.95 | Improved |
| Set γ to 0.5 rather than 1.0 (ablation) | PS-CD · CIFAR-10 (32 × 32) conditional | FID: 29.78 -> 35.02 | Hurt |
| Set γ to -0.5 rather than 1.0 (ablation) | PS-CD · CelebA (64 × 64) | FID: 24.76 -> 20.35 | Improved |
| Train with PS-CD rather than CD | Gaussian EBM · Gaussian contamination ratio 0.2 | KL divergence: 0.3869 -> 6.4e-6 | Improved |
| Increase contamination ratio from 0.01 to 0.3 | Contrastive Divergence · Gaussian contamination data | KL divergence: 0.0067 -> 0.5438 | Hurt |
| Use γ=2.0 rather than γ=0.5 (ablation) | PS-CD · Gaussian contamination ratio 0.3 | KL divergence: 0.5429 -> 0.00029 | Improved |
| Vary γ across PS-CD objectives (ablation) | PS-CD · Mis-specified 1-D Gaussian mixture fitted with a single Gaussian | entropy | Improved |
| Use PS-CD rather than CD | EBM · 2-D synthetic datasets | MMD | Mixed |
Closest papers by meaning
- Near-Optimality of Contrastive Divergence Algorithms (NeurIPS 2024 · similarity 0.87)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.85)
- Bridging Mini-Batch and Asymptotic Analysis in Contrastive Learning: From InfoNCE to Kernel-Based Losses (ICML 2024 · similarity 0.84)
- Joint Learning of Energy-based Models and their Partition Function (ICML 2025 · similarity 0.83)
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.83)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · 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.
