Thursday

The expected data rates from the next generation of astronomical facilities have resulted in the increased use of deep learning for survey analysis. However, such approaches typically require large training sets and furthermore produce models with millions of learnable parameters. This can be problematic in terms of both storage requirements and computational cost at deployment. In this talk we present a variational inference (VI) approach to radio galaxy classification with convolutional neural networks, based on the so-called "bayes-by-backprop" methodology. Our results show that this method naturally allows for large deep learning models to be substantially weight-pruned in order to minimise model size without compromising model performance. We will also discuss possible biases introduced by the cold-posterior problem seen in other applications of VI-based deep learning in the context of radio astronomy classification, where it is of particular relevance as data augmentation is commonly used to increase the size of small training sets. We will consider the generalisation of Bayesian neural networks in radio astronomy under domain shift or model misspecification and discuss whether the use of an additional variance term in the posterior might be justified in such circumstances.