Source-linked AI summary
Matching matched filtering with deep networks in gravitational-wave astronomy
Hunter Gabbard, Michael Williams, Fergus Hayes, Chris Messenger
TL;DR
Matched filtering is computationally expensive for gravitational-wave searches as detector sensitivity improves at low frequencies. This paper trains a deep CNN on whitened BBH time series in Gaussian noise and finds sensitivity closely matching matched filtering, while noting that real non-Gaussian noise remains outside the demonstrated setting.
Problem
Gravitational-wave search pipelines are computationally expensive, motivating a low-latency signal-versus-noise method that can match matched-filtering sensitivity.
Method
The authors train a deep convolutional neural network using raw whitened gravitational-wave time series and compare its classification output with matched filtering on simulated BBH signals and Gaussian noise.
Results
The CNN closely matches matched-filtering sensitivity across explored false alarm probabilities, with higher sensitivity at low SNR and high false alarm probability but marginally lower sensitivity at high SNR and low false alarm probability.
Takeaways & Limitations
A deep CNN can reproduce matched-filtering performance for BBH detection in Gaussian noise using pre-trained time-series analysis suitable for low-latency application.
Takeaways & Limitations
The analysis uses Gaussian noise, whereas real gravitational-wave searches are affected by non-Gaussian noise artifacts and require modified ranking statistics and data-quality excision.
Abstract
from arXiv · showhide
We report on the construction of a deep convolutional neural network that can reproduce the sensitivity of a matched-filtering search for binary black hole gravitational-wave signals. The standard method for the detection of well modeled transient gravitational-wave signals is matched filtering. However, the computational cost of such searches in low latency will grow dramatically as the low frequency sensitivity of gravitational-wave detectors improves. Convolutional neural networks provide a highly computationally efficient method for signal identification in which the majority of calculations are performed prior to data taking during a training process. We use only whitened time series of measured gravitational-wave strain as an input, and we train and test on simulated binary black hole signals in synthetic Gaussian noise representative of Advanced LIGO sensitivity. We show that our network can classify signal from noise with a performance that emulates that of match filtering applied to the same datasets when considering the sensitivity defined by Reciever-Operator characteristics.