A debiasing framework for deep learning applied to the morphological classification of galaxies.

Abstract

In order to train deep learning models, usually a large amount of correctly annotated data is needed. Depending on the data domain, the task of correctly annotating data can prove to be difficult, as in many cases the ground truth of the data is not obtainable. This is true for numerous problems within the astronomy domain, one of these being the morphological classification of galaxies. The aforementioned means that astronomers are forced to rely on an estimate of the ground truth, often generated by human annotators. The problem with this is that human generated labels have been shown to contain biases related to the quality of the data being labeled, such as image resolution. This type of bias is a consequence of the quality of the data, that is, it is independent of the annotators, meaning that even datasets annotated by experts can be affected by this type of bias. In this work, we show that deep learning models trained on biased data learn the bias contained in the data, transferring the bias to its predictions. We also propose a framework to train deep learning models, that allows us to obtain unbiased models even when training on biased data. We test our framework by training a classification model on images of morphologically classified galaxies from Galaxy Zoo 2 and show that we are able to diminish the bias in the data.