Fanaroff-Riley classification of radio galaxies using group-equivariant convolutional neural networks

TL;DR

This paper applies group-equivariant CNNs to radio galaxy classification, leveraging symmetry properties to improve performance and uncertainty estimation, addressing the challenge of orientation invariance in astronomical image analysis.

Contribution

First application of group-equivariant CNNs to radio galaxy classification, demonstrating improved performance and uncertainty estimation by incorporating Euclidean group symmetries.

Findings

Equivariant models outperform conventional CNNs in classification accuracy.

E(2)-equivariant models reduce confidence variation with rotation.

D16 equivariant model achieves the best test performance.

Abstract

Weight sharing in convolutional neural networks (CNNs) ensures that their feature maps will be translation-equivariant. However, although conventional convolutions are equivariant to translation, they are not equivariant to other isometries of the input image data, such as rotation and reflection. For the classification of astronomical objects such as radio galaxies, which are expected statistically to be globally orientation invariant, this lack of dihedral equivariance means that a conventional CNN must learn explicitly to classify all rotated versions of a particular type of object individually. In this work we present the first application of group-equivariant convolutional neural networks to radio galaxy classification and explore their potential for reducing intra-class variability by preserving equivariance for the Euclidean group E(2), containing translations, rotations and reflections. For the radio galaxy classification problem considered here, we find that classification performance is modestly improved by the use of both cyclic and dihedral models without additional hyper-parameter tuning, and that a D16 equivariant model provides the best test performance. We use the Monte Carlo Dropout method as a Bayesian approximation to recover epistemic uncertainty as a function of image orientation and show that E(2)-equivariant models are able to reduce variations in model confidence as a function of rotation.