Improving Background Rejection with Normalizing Flows in Anomalous Signal Searches within Derived Spaces

After the discovery of the Higgs Boson, researchers have been attempting to find new particles that could explain the phenomena that cannot be explained by the particles in the Standard Model of physics, such as dark matter. Deep Learning approaches are a popular method to try to find anomalous events. We propose the Quasi Anomalous Knowledge (QUAK) semi-supervised deep learning approach which develops a 2D space (QUAK space) where the axes capture how signal-like or background-like an event is. The axes are based on Normalizing Flows (NF) where one is trained on simulated background and the other is trained on postulated beyond standard model scenarios – these give the likeness to background or signal for an event. The background, signals, and anomalies live in different areas of this QUAK space. By training an NF, a generative model, on the QUAK space, the NF can create a Probability Density Function which estimates the distribution of events in the space. The NF is conditionally trained on the invariant jet mass for different mass bins to capture the mass dependence of events. An interpolation scheme across generated mass bins allows for a background subtraction which creates a flat distribution where resonances are more apparent.