New Objective Improves Photometric Redshift Estimation
Analysis
This paper introduces Starkindler, a novel training objective for photometric redshift estimation that explicitly accounts for aleatoric uncertainty (observational errors). This is a significant contribution because existing methods often neglect these uncertainties, leading to less accurate and less reliable redshift estimates. The paper demonstrates improvements in accuracy, calibration, and outlier rate compared to existing methods, highlighting the importance of considering aleatoric uncertainty. The use of a simple CNN and SDSS data makes the approach accessible and the ablation study provides strong evidence for the effectiveness of the proposed objective.
Key Takeaways
- •Starkindler is a new training objective for photometric redshift estimation.
- •It explicitly incorporates observational errors (aleatoric uncertainty).
- •It improves accuracy, calibration, and reduces outlier rates compared to existing methods.
- •The approach is validated using SDSS data and a simple CNN.
- •The method provides interpretable uncertainty estimates.
“Starkindler provides uncertainty estimates that are regularised by aleatoric uncertainty, and is designed to be more interpretable.”