Astrophysics informed neural networks for signal detection

In transient astrophysics, the Universe itself is the Laboratory, individual cosmic events are probabilistic and cannot be repeated. Therefore, reliable and efficient signal detection methods are needed. Analysis of astrophysics data through machine learning approaches have skyrocketed in recent years in many subfields of astrophysics including gravitational wave science but the applied methods often lack interpretability.

High quality gravitational-wave data’s interpretability is inherently fingerprinted in the detected waveform itself. We make a key observation on the relationship between deep learning and matched filtering, the standard method for gravitational-wave detection: matched filtering with a collection of templates is formally equivalent to a particular neural network suggesting new perspectives on the role of deep learning in gravitational wave detection. A neural network thus can be constructed analytically to exactly implement matched filtering and can be further trained on data or boosted with additional complexity for improved performance.