Abstract:
    Obtaining accurate and well-calibrated photometric redshift (photo-z) probability densities for galaxies without spectroscopic measurements remains a challenge. Although deep learning methods hold great promise in fulfilling the need for photo-z estimation for future surveys, they usually lack controllability, suffer from miscalibration and thus may not be directly applicable to astrophysical and cosmological analysis. I will present our novel photo-z estimation method “CLAP” that aims to resolve this issue. In a nutshell, CLAP leverages deep learning to extract redshift information from multi-band galaxy images, and exploits adaptive k-nearest neighbors to predict photo-z probability density estimates. As will be demonstrated, CLAP achieves substantially better calibration of probability density estimates compared to benchmark methods, and in the meantime retains high accuracy and computational efficiency. With reference to CLAP, the shortcomings of conventional deep learning methods will also be pointed out. In a more general sense, this work is one of our attempts to promote reliable “AI+astronomical” research in the context of big astronomical data.