Current data analysis pipelines are limited by the extreme computational costs of template-based matched-ﬁltering methods and thus are unable to scale to all types of sources. We introduced Deep Filtering, a new method for end-to-end time-series signal processing which combines two deep convolutional neural networks to rapidly detect and estimate parameters of signals much weaker than the background noise.
This is the foundation research on deep learning for gravitational wave detection, and can be rightly regarded as the textbook reference that established for the first time the power of deep learning to outperform other gravitational wave analysis methods in terms of both accuracy and speed.
We show for the very first time that deep convolutional neural networks can match the sensitivity of matched-filtering searches for detecting signals in noisy time-series data. We also prove that deep learning can identify new classes of signals beyond the scope of existing methodologies.
We applied this method for gravitational wave analysis specifically for mergers of black holes and demonstrated that it significantly outperforms conventional machine learning techniques, is far more efficient than matched-filtering allowing real-time processing of raw big data with minimal resources and extends the range of gravitational waves that can be detected by advanced LIGO.
A follow-up article shows for the first time that deep learning can detect true gravitational wave signals in real LIGO data. Here we show that neural networks can be used in realistic detection scenarios, and can learn to adapt to the non-Gaussian and non-stationary behavior of real LIGO data. To demonstrate this novel detection method, we showed that we can correctly identify and estimate the properties of the first gravitational wave detection (GW150914) in this article
Project Members: Daniel George, Eliu Huerta