Domain-Adaptive Neural Posterior Estimation for Strong Gravitational Lens Analysis

TL;DR

This work performs the first study of the efficacy of NPE in combination with unsupervised domain adaptation (UDA) and finds that combining UDA and NPE improves the accuracy of the inference by 1-2 orders of magnitude and significantly improves the posterior coverage over an NPE model without UDA.

Abstract

Modeling strong gravitational lenses is prohibitively expensive for modern and next-generation cosmic survey data. Neural posterior estimation (NPE), a simulation-based inference (SBI) approach, has been studied as an avenue for efficient analysis of strong lensing data. However, NPE has not been demonstrated to perform well on out-of-domain target data -- e.g., when trained on simulated data and then applied to real, observational data. In this work, we perform the first study of the efficacy of NPE in combination with unsupervised domain adaptation (UDA). The source domain is noiseless, and the target domain has noise mimicking modern cosmology surveys. We find that combining UDA and NPE improves the accuracy of the inference by 1-2 orders of magnitude and significantly improves the posterior coverage over an NPE model without UDA. We anticipate that this combination of approaches will help enable future applications of NPE models to real observational data.

Figures (2)

• Figure 1: (a): Two example images from the source (without noise; left) and target (with noise; right) domains. (b): Feature spaces of the embedding network when models are applied to the source (points are filled circles; all points encompassed by a blue circle) and target (points are filled triangles; all points encompassed by an orange circle) domain data for the NPE-only (left) and NPE-UDA (right) models, respectively. (c): Residuals on the five lens parameters ($\theta_{\mathrm{E}}$, $x$, $y$, $e_{\mathrm{l},1}$, $e_{\mathrm{l},2}$) for the NPE-only model applied to target data (orange), the NPE-UDA model applied to target data (blue), and the NPE-UDA model applied to source data (pink). Contours show the 68th- and 95th-percentile confidence regions, and the dashed lines show zero residuals. (d): Posterior coverage on the five lens parameters for the NPE-only model for the NPE-UDA model applied to target data (dashed, color), the NPE-only model applied to target data (solid, color), and the NPE-UDA model applied to source data (solid, black). The boundary between underconfident (upper) and overconfident (lower) is marked by a dotted gray line.
• Figure 2: Latent space of the embedding network when NPE is applied to the source and target domain test set data for the NPE-only (left) and NPE-UDA (right) models, respectively. This is applied to parameters $x$(a), $y$(b), $e_{\mathrm{l},1}$(c), $e_{\mathrm{l},2}$(d).