Orateur
Description
Ever since X-ray observatories have started to probe the Universe, one of the objectives has been to classify the observed sources. Indeed, the X-ray sky is extremely diverse, with emission coming from stellar atmospheres, to super-massive black holes, up to galaxy clusters. It remains challenging to conduct extensive studies of all the detected sources due to the considerable volume of data that needs to be analyzed. For instance, most of the classification work has been supervised, meaning sources have been sorted in human-decided categories, based on their properties. These categories can be somewhat arbitrary, and will also be challenged by future missions probing new, never-before-seen depths of the Universe.
The work presented here is an attempt at unsupervised classification of the entire XMM-Newton archive. For this purpose, we apply a Variational Auto-Encoder (VAE) architecture to the EPIC pn spectra of each detection. Studying the latent space of the VAE allows us to dramatically reduce the dimensionality of the classification problem, down to the (relatively low) dimension of the latent space. We can then analyse how these objects scatter in latent space, how they tend to cluster with similar objects, and how they evolve over time through latent trajectories. While this is still a work in progress, we prove the potential of this method in both unsupervised classification, as well as possible applications in similarity search, or in outlier detection.
