Opening the Black Box of Neural Nets: Case Studies in Stop/top Discrimination

Abstract

In this thesis we examine several different approaches for deducing observable signals of new physics. First, we outline a technique to extract useful information about the pattern a neural network has learned to recognize, and use it to deduce novel correlations to discriminate top and scalar top quark decays. Then, we propose a model of dark matter which is able to evade cosmological constraints to simultaneously manifest the complicated self-interactions of “Double Disk Dark Matter” and interactions with the Standard Model strong enough to potentially show up in near-future detectors. It achieves this by decoupling the dark and visible sectors through late-stage asymmetric reheating. Finally, we survey possible dark matter decays which could result in the observed galactic X-ray excess at 3.5 keV and conclude that, besides the common explanation in terms of sterile neutrinos, there is also a natural explanation in terms of the decay of heavy scalar moduli. In this case, the energy of the X-ray line implies a scale of supersymmetry breaking at or below 1000 TeV.