Bignone Lucas
Argentina – AFE (UBA/CONICET)
Disentangling galaxy morphologies with generative neural networks
Galaxy morphologies have been typically studied using a variety of parametric and non-parametric methods, as well as visual classification. Most of these methods have been devised to study low and intermediate redshift galaxies in the high surface brightness regime. As new instruments such as JWST, Euclid and Vera Rubin reveal galaxies in previously unexplored regions, it becomes very important that we revise the methods and taxonomies we use to classify galaxies so that we can properly link their visual appearance to their physical properties and origin. In recent years, generative neural network models, including Generative Adversarial Networks (GANs), autoencoders, and diffusion models, have made a significant impact across multiple disciplines. These advanced AI algorithms enable the generation of realistic and diverse data samples, fostering advancements in tasks like image synthesis, data augmentation, and style transfer, thereby pushing the boundaries of visual understanding. The power of these models lies in their ability to efficiently capture and organize visual information within images in interpretable and disentangled ways. In this work, we propose an unsupervised classification method for galaxy morphology based on GANs deep neural networks. Our model’s flexibility allows it to adapt to various datasets and domains without assuming a predefined classification taxonomy. We also implement methods to automatically extract semantic meaning from the neural network’s internal representations of galaxies in the latent space, enhancing our understanding of the network’s classifications. Our approach has been tested on diverse datasets, including mock simulated galaxy images from the EAGLE simulation, deep images from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP), and high-resolution images from The Dark Energy Camera Legacy Survey (DECaLS). The remarkable ability of our method to extract meaningful information from raw galaxy images proves valuable in situations where labeled datasets are impractical or existing methods are not applicable. These neural networks’ capacity to uncover hidden patterns, structures, and relationships within data holds great potential for driving innovation and discovery.