Empirical Study of MC-Dropout in Various Astronomical Observing Conditions

Noble Kennamer, Alex Ihler, David Kirkby; The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 17-20

Abstract


The analysis of large astronomical surveys increasingly incorporates machine learning models to handle a diverse set of tasks. It is important for the scientific analysis of these surveys that the uncertainty of the models be well understood and the predictions properly calibrated. Here we present an empirical study of MC-Dropout for a core prediction problem in astronomy emphasizing how the modeled uncertainty is influenced by changes in observing conditions. We will show that while MC-Dropout results in improved accuracy and better calibrated predictions there is still an underestimation of uncertainty that needs to be addressed.

Related Material


[pdf]
[bibtex]
@InProceedings{Kennamer_2019_CVPR_Workshops,
author = {Kennamer, Noble and Ihler, Alex and Kirkby, David},
title = {Empirical Study of MC-Dropout in Various Astronomical Observing Conditions},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}