Seminar Details
2025-02-18 (14h) : Deep Learning Mass Mapping with Conformal Predictions
At Euler building (room A.002)
Organized by Mathematical Engineering
Speaker :
Hubert Leterme (Ensicaen, CEA Paris-Saclay)
Abstract :
In this talk, I will introduce a plug-and-play (PnP) approach for mapping the distribution of dark matter in the sky using weak gravitational lensing and noisy shear measurements. This method aims to deliver accurate and efficient mass estimates without the need to retrain deep learning models for each new galaxy survey or sky region. Instead, a single model is trained on simulated convergence maps with Gaussian white noise corruption. To enhance reliability, we incorporate a distribution-free uncertainty quantification (UQ) method, conformalized quantile regression (CQR), into the mass mapping framework. Leveraging a simulation-derived calibration set, CQR provides rigorous coverage guarantees, independent of any specific prior data distribution. We benchmark this PnP approach against existing mass mapping methods, including Kaiser-Squires, Wiener filtering, MCALens, and DeepMass. Our results demonstrate that while miscoverage rates, after calibration with CQR, remain consistent across methods, the choice of mass mapping technique significantly impacts the size of the resulting error bars.