High-dimensional inference for the $γ$-ray sky with differentiable programming

We motivate the use of differentiable probabilistic programming techniques in order to account for the large model-space inherent to astrophysical γ-ray analyses. Targeting the longstanding Galactic Center γ-ray Excess (GCE) puzzle, we construct differentiable forward model and likelihood that make liberal use of GPU acceleration and vectorization in order to simultaneously account for a continuum of possible spatial morphologies consistent with the GCE emission in a fully probabilistic manner. Our setup allows for efficient inference over the large model space using variational methods. Beyond application to γ-ray data, a goal of this work is to showcase how differentiable probabilistic programming can be used as a tool to enable flexible analyses of astrophysical datasets.