Image Details

Choose export citation format:

Improving Generalization and Uncertainty Quantification of Photometric Redshift Models

  • Authors: Jonathan Soriano, Tuan Do, Srinath Saikrishnan, Vikram Seenivasan, Bernie Boscoe, Jack Singal, Evan Jones

Jonathan Soriano et al 2026 The Astronomical Journal 171 .

  • Provider: AAS Journals

Caption: Figure 13.

Comparison of statistical coverage per tomographic bin before and after applying split conformal prediction to BNN predictions. Ideally, 68% of evaluated galaxies will have the ground-truth redshift within their 68% prediction interval. A percentage of galaxies above this target level indicates galaxies are “overcovered” or have wide prediction intervals. A percentage of galaxies below this target level indicate “undercovered” or narrow prediction intervals. Before split conformal prediction, BNN-evaluated galaxies from the GalaxiesML and TransferZ samples produce wide prediction intervals at low redshift and tend to become narrow at higher redshifts. After split conformal prediction, the desired statistical coverage level is achieved at the full redshift range for GalaxiesML; however, only BNN-2 achieves accurate coverage on the TransferZ sample.

Other Images in This Article

Show More

Copyright and Terms & Conditions

Additional terms of reuse