Variational views for self-supervised learning in radio astronomy

TL;DR

Results indicate that generative and contrastive approaches are complementary, and point toward disentanglement-aware self-supervised learning as a promising direction for future radio astronomy surveys.

Abstract

Modern astronomical surveys are producing progressively larger and more complex datasets, making traditional supervised approaches that rely on extensive labelled catalogues increasingly difficult. Consequently, pre-training using self-supervised learning (SSL), which offers a scalable route by extracting structure directly from unlabelled images, is becoming attractive for many downstream applications. In this work we consider the use of coupled self-supervised representation learning approaches for radio galaxy morphology pre-training. In order to account for the more nuanced variations in radio galaxy morphology than are typically included in the augmented views of view-based SSL algorithms, we use a pre-trained Variational Autoencoder (VAE) to generate views for training a larger view-based self-supervised model. To do this, a $β$-VAE was trained on the Radio Galaxy Zoo (RGZ) dataset, where moderate regularization ($β= 2.3$) was found to provide a good balance between reconstruction quality and disentanglement of generative factors such as source multiplicity and lobe asymmetry. An analysis of the $β$-VAE reveals that Fanaroff-Riley class identity manifests as a continuous transition across the latent space, rather than being associated to a single discrete dimension. $β$-VAE reconstructions were then incorporated as generative augmentations within a view-based SSL pipeline. Our experiments show that combining these generative views with standard image augmentations improves downstream classification performance, and we present ablation studies clarifying the relative contribution of each augmentation type. These results indicate that generative and contrastive approaches are complementary, and point toward disentanglement-aware self-supervised learning as a promising direction for future radio astronomy surveys.

Figures (12)

• Figure 1: Schematic of the BYOL framework with online and target networks.
• Figure 2: Example radio galaxy cutouts from the MiraBest dataset, showing confident samples of (a) FR I, (b) FR II, and (c) Hybrid morphologies.
• Figure 3: Example radio galaxy cutouts from the RGZ dataset.
• Figure 4: Diagnostic plot of Regularization Strength ($\beta$) vs. Latent Capacity for the RGZ dataset. The balance point where the curves intersect aligns with the empirically chosen $\beta=2.3$.
• Figure 5: Reconstructions of a MiraBest radio source with a $\beta$-VAE ($\beta = 1.22$). The original is shown at left in each row, followed by multiple stochastic reconstructions obtained by sampling the latent code with increasing variance: (A) $\mathcal{N}(0,0.1)$, (B) $\mathcal{N}(0,0.5)$, (C) $\mathcal{N}(0,1.0)$.