Training strategies for deep learning gravitational-wave searches

Compact binary systems emit gravitational radiation which is potentially detectable by current Earth bound detectors. Extracting these signals from the instruments' background noise a complex problem and computational cost of most searches depends on complexity source model. Deep learning may be capable finding where algorithms hit limits. Here we restrict our analysis to non-spinning black holes systematically test different strategies training data presented networks. To assess impact strategies, re-analyze first published networks directly compare them an equivalent matched-filter search. We find that deep can generalize low signal-to-noise ratio (SNR) high SNR ones but not vice versa. As such, it beneficial provide during training, fastest convergence achieved when samples are provided early on. During testing found sometimes unable recover any false alarm probability $<10^{-3}$ required. resolve this restriction applying modification call unbounded Softmax replacement (USR) after training. With alteration machine search retains $\geq 91.5\%$ sensitivity down false-alarm rate 1 per month.