Research map / Probabilistic and scientific modeling / Variational and generative sampling
``Noisier'’ Noise Contrastive Estimation is (Almost) Maximum Likelihood
Peiyu Yu, Dinghuai Zhang, Hengzhi He, Xiaojian Ma, Sirui Xie, RUIYAO MIAO, Yifan Lu, Yasi Zhang, Deqian Kong, Ruiqi Gao, Jianwen Xie, Guang Cheng and 1 more
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 |
|---|---|---|---|
| scale noise magnitude M from 1 to 100 (ablation) | LEBM on CelebAHQ (nz=512) · CelebAHQ FID | FID: 118.84 -> 101.71 | Improved |
| add multi-stage ratio estimation (K=3 vs K=1) (ablation) | N2CE-LEBM · CelebAHQ FID | FID: 101.71 -> 95.66 | Improved |
| N2CE instead of MLE with short-run Langevin (ablation) | LEBM · SVHN FID | FID: 32.74 -> 25.63 | Improved |
| N2CE (M=100) instead of NCE (ablation) | DxMI reward learning · CIFAR-10 FID | FID: 3.93 -> 2.99 | Improved |
| N2CE (M=100) instead of NCE (ablation) | DxMI reward learning · ImageNet64x64 FID | FID: 2.69 -> 2.23 | Improved |
| N2CE (M=50) instead of NCE (M=1) critic (ablation) | SiD · CIFAR-10 unconditional FID | FID: 1.53 -> 1.45 | Improved |
| N2CE (M=50) instead of NCE (M=1) critic (ablation) | SiD · ImageNet64x64 FID | FID: 1.28 -> 1.11 | Improved |
| N2CE instead of NCE in DAMC prior (ablation) | DAMC anomaly detection · MNIST held-out digit 9 AUPRC | AUPRC: 0.502 -> 0.854 | Improved |
| N2CE instead of NCE in DAMC prior (ablation) | DAMC anomaly detection · MNIST held-out digit 7 AUPRC | AUPRC: 0.605 -> 0.845 | Mixed |
| N2CE (M=100,K=6) instead of MLE-LEBM with LD sampler (ablation) | offline BBO · Design-Bench mean score | mean score: 0.562 -> 0.748 | Improved |
| N2CE (M=100,K=6) instead of NCE (M=1,K=6) with SVGD (ablation) | offline BBO · Design-Bench mean score | mean score: 0.770 -> 0.827 | Improved |
| reduce stages from K=6 to K=1 in N2CE (ablation) | offline BBO · Design-Bench mean score | mean score: 0.827 -> 0.764 | Hurt |
Closest papers by meaning
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.86)
- Information Theoretic Learning for Diffusion Models with Warm Start (NeurIPS 2025 · similarity 0.86)
- Classification Diffusion Models: Revitalizing Density Ratio Estimation (NeurIPS 2024 · similarity 0.86)
- Noise Conditional Variational Score Distillation (ICML 2025 · similarity 0.85)
- Your Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training (NeurIPS 2024 · similarity 0.85)
- A Unified View on Learning Unnormalized Distributions via Noise-Contrastive Estimation (ICML 2025 · 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.
