Explaining deep learning of galaxy morphology with saliency mapping

Prabh Bhambra,Benjamin Joachimi,Ofer Lahav

doi:10.1093/mnras/stac368

Prabh Bhambra, Benjamin Joachimi + Show 1 more

Open Access

https://doi.org/10.1093/mnras/stac368

Copy DOI

Abstract

ABSTRACT We successfully demonstrate the use of explainable artificial intelligence (XAI) techniques on astronomical data sets in the context of measuring galactic bar lengths. The method consists of training convolutional neural networks on human classified data from Galaxy Zoo in order to predict general galaxy morphologies, and then using SmoothGrad (a saliency mapping technique) to extract the bar for measurement by a bespoke algorithm. We contrast this to another method of using a convolutional neural network to directly predict galaxy bar lengths. These methods achieved correlation coefficients of 0.76 and 0.59, and root mean squared errors of 1.69 and 2.10 respective to human measurements. We conclude that XAI methods outperform conventional deep learning in this case, which could be reasonably explained by the larger data sets available when training the models. We suggest that our XAI method can be used to extract other galactic features (such as the bulge-to-disc ratio) without needing to collect new data sets or train new models. We also suggest that these techniques can be used to refine deep learning models as well as identify and eliminate bias within training data sets.

Highlights

From autonomous vehicles to targeted advertising, machine learning (ML) algorithms have become more and more ubiquitous in society and a greater emphasis is being put on understanding the decisions they make
This paper aims to use XAI methods to generalise ML models trained on astronomical datasets
We aim to show that XAI methods can be used to extract new information, which has never been presented to the convolutional neural networks (CNNs), from its classifications

Summary

Introduction

From autonomous vehicles to targeted advertising, machine learning (ML) algorithms have become more and more ubiquitous in society and a greater emphasis is being put on understanding the decisions they make. Algorithms themselves vary in their interpretability, with the likes of decisions trees tending towards being more interpretable, while neural networks tend towards being some of the least interpretable (Došilović et al 2018). Despite their opacity, neural networks are widely used due to their relatively good performance compared to other ML algorithms. Due to the richness of the data, the Galaxy Zoo project was chosen (Willett et al 2013b) To our knowledge this is the first attempt at quantitatively using saliency mapping for image processing in astronomy, there are recent cases of them being used qualitatively for spectral analysis (Peruzzi et al 2021) and image processing (Villanueva-Domingo & Villaescusa-Navarro 2021). This paper intends to act as a proof of concept that these techniques are effective when applied to this field, and that they can be used to infer quantitative measurements

Objectives

Methods

Conclusion