Research map / Probabilistic and scientific modeling / Variational and generative sampling
Cross-Fitting for Neural Posterior Estimation
Jeffrey Regier
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes cross-fitting for neural posterior estimation in simulation-based inference; the abstract is truncated so details beyond the problem framing are unknown.
Problem
Neural Posterior Estimation trains an amortized conditional density estimator from single (latent, observation) pairs, so posterior variance must be inferred across pairs; a flexible variational distribution can collapse or be poorly calibrated in this setting.
Approach
Applies cross-fitting (a sample-splitting technique from semiparametric statistics) to NPE training; the abstract only partially states the idea, so the specific mechanism is not reported here.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace standard NPE with CF-NPE (M=1) | NPE with Gaussian encoder · Gaussian-Gaussian (d=200) | test log-likelihood: -141.30 -> -133.92 | Improved |
| replace standard NPE with CF-NPE (M=1) | NPE with NSF · Funnel (d=200) | test log-likelihood: -58.98 -> 38.09 | Improved |
| replace standard NPE with CF-NPE | NPE with NSF · Strong-lensing source reconstruction | test log-likelihood (median): 2.59 -> 60.05 | Improved |
| fit stage 2 to out-of-fold instead of in-fold residuals (cross-fitting) (ablation) | CF-NPE · Gaussian-Gaussian (d=200), Gaussian | chi-square d: 1.150 -> 0.993 | Improved |
| add K-fold cross-fitting and replicates to a single stage-1 mean fit (ablation) | CF-NPE · Gaussian-Gaussian (d=200), Gaussian | test log-likelihood: -136.01 -> -128.42 | Improved |
| increase number of folds K from 2 to 16 (ablation) | CF-NPE (M=3) · Gaussian-Gaussian (d=200) | test log-likelihood: -132.93 -> -128.18 | Improved |
| increase stage-1 ensemble size M from 1 to 10 (ablation) | CF-NPE · Gaussian-Gaussian (d=200), Gaussian | test log-likelihood: -133.92 -> -126.41 | Improved |
| use mixture of NPE networks (deep-ensemble style) instead of single NPE (ablation) | Std-NPE · Gaussian-Gaussian (d=200), Gaussian | chi-square d: LL -141.30 -> -133.28 but chi2 1.095 -> 0.875 | Mixed |
| form stage-2 mixture (CF-NPE-mixture) on top of CF-NPE (ablation) | CF-NPE, M=3 · Gaussian-Gaussian (d=200) | test log-likelihood: -128.42 -> -128.41 | No clear effect |
| form stage-2 mixture on top of CF-NPE with NSF (ablation) | CF-NPE, M=3 · Strong-lensing source reconstruction | test log-likelihood (median): 83.21 -> 88.08 | Improved |
| oversize the stage-2 NSF (ablation) | CF-NPE · Gaussian-Gaussian (d=200) | test log-likelihood | Hurt |
| apply temperature scaling to standard NPE (ablation) | Std-NPE · Gaussian-Gaussian (d=200) | test log-likelihood | No clear effect |
Experimental setup
Builds on: neural posterior estimation
Closest papers by meaning
- Neural Posterior Estimation with Latent Basis Expansions (ICLR 2026 · similarity 0.89)
- Sequential Neural Score Estimation: Likelihood-Free Inference with Conditional Score Based Diffusion Models (ICML 2024 · similarity 0.86)
- CoLT: The conditional localization test for assessing the accuracy of neural posterior estimates (NeurIPS 2025 · similarity 0.85)
- Variational Inference with Coverage Guarantees in Simulation-Based Inference (ICML 2024 · similarity 0.85)
- Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors (AAAI 2025 · similarity 0.84)
- Globally Convergent Variational Inference (NeurIPS 2024 · 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.
