Conservative neural posterior estimation via distributionally robust training
Abstract
Simulation-based inference with neural posterior estimation (NPE) often yields overconfident and unreliable posteriors under limited simulation budgets. To address this, we propose DRO-NPE, a distributionally robust approach that replaces the standard NPE objective with a worst-case loss over a Wasserstein ambiguity set. We introduce KL-based metrics for miscoverage and miscalibration, and use these to show that the DRO-NPE objective controls overfitting and reduces posterior overconfidence. Our method is tractable, parallelisable, and readily integrates with standard normalising flows. Across benchmark SBI tasks, DRO-NPE consistently improves coverage and calibration, while narrowing the gap between empirical and population NPE loss, leading to more reliable inference in low-simulation regimes.
Related
- Joint Neural Posterior Estimation and Model Selection for Chemical Reaction Networks
- Cross-Fitting for Neural Posterior Estimation
- Neural Posterior Estimation for Tomographic Weak Lensing Mass Mapping
- Amortized Data Borrowing with Exchangeability-Aware Neural Posterior Estimation
- Neural Posterior Estimation for Spatial Individual-Level Epidemic Models
BibTeX
@misc{laplante2026conservative,
title = {{Conservative neural posterior estimation via distributionally robust training}},
author = {William Laplante and Yuga Hikida and Charita Dellaporta and François-Xavier Briol and Ayush Bharti},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2605.28516},
eprint = {2605.28516},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.28516},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Conservative neural posterior estimation via distributionally robust training
AU - William Laplante
AU - Yuga Hikida
AU - Charita Dellaporta
AU - François-Xavier Briol
AU - Ayush Bharti
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2605.28516
UR - https://arxiv.org/abs/2605.28516
AB - Simulation-based inference with neural posterior estimation (NPE) often yields overconfident and unreliable posteriors under limited simulation budgets. To address this, we propose DRO-NPE, a distributionally robust approach that replaces the standard NPE objective with a worst-case loss over a Wasserstein ambiguity set. We introduce KL-based metrics for miscoverage and miscalibration, and use these to show that the DRO-NPE objective controls overfitting and reduces posterior overconfidence. Our method is tractable, parallelisable, and readily integrates with standard normalising flows. Across benchmark SBI tasks, DRO-NPE consistently improves coverage and calibration, while narrowing the gap between empirical and population NPE loss, leading to more reliable inference in low-simulation regimes.
ER -