Convolutional neural networks: A magic bullet for gravitational-wave detection?

Gebhard, Timothy D.; Kilbertus, Niki; Harry, G. M.; Schölkopf, Bernhard

doi:10.1103/physrevd.100.063015

Cited by 138 publications

(103 citation statements)

References 71 publications

Supporting

Mentioning

102

Contrasting

Order By: Relevance

“…More recently DNNs have started to draw attention in the field of gravitational-wave searches (i) as a classifier for non-Gaussian detector transients (glitches) [24][25][26][27], (ii) as a search method for unmodeled burst signals [28,29] in time-frequency images produced by coherent WaveBurst [30], and (iii) as a direct detection method for black-hole merger signals in gravitational-wave strain data [31][32][33][34][35][36].…”

Section: Introductionmentioning

confidence: 99%

Deep-learning continuous gravitational waves

et al. 2019

View full text Add to dashboard Cite

We present a first proof-of-principle study for using deep neural networks (DNNs) as a novel search method for continuous gravitational waves (CWs) from unknown spinning neutron stars. The sensitivity of current wide-parameter-space CW searches is limited by the available computing power, which makes neural networks an interesting alternative to investigate, as they are extremely fast once trained and have recently been shown to rival the sensitivity of matched filtering for black-hole merger signals [D. George and E. A. Huerta, Phys. Rev. D 97, 044039 (2018); H. Gabbard, M. Williams, F. Hayes, and C. Messenger, Phys. Rev. Lett. 120, 141103 (2018)]. We train a convolutional neural network with residual (shortcut) connections and compare its detection power to that of a fully coherent matchedfiltering search using the WEAVE pipeline [K. Wette, S. Walsh, R. Prix, and M. A. Papa, Phys. Rev. D 97, 123016 (2018)]. As test benchmarks we consider two types of all-sky searches over the frequency range from 20 to 1000 Hz: an "easy" search using T ¼ 10 5 s of data, and a "harder" search using T ¼ 10 6 s. The detection probability p det is measured on a signal population for which matched filtering achieves p det ¼ 90% in Gaussian noise. In the easiest test case (T ¼ 10 5 s at 20 Hz) the DNN achieves p det ∼ 88%, corresponding to a loss in sensitivity depth of ∼5% versus coherent matched filtering. However, at higher frequencies and for longer observation times the DNN detection power decreases, until p det ∼ 13% and a loss of ∼66% in sensitivity depth in the hardest case (T ¼ 10 6 s at 1000 Hz). We study the DNN generalization ability by testing on signals of different frequencies, spindowns and signal strengths than they were trained on. We observe excellent generalization: only five networks, each trained at a different frequency, would be able to cover the whole frequency range of the search.

show abstract

Section: Introductionmentioning

confidence: 99%

Deep-learning continuous gravitational waves

et al. 2019

View full text Add to dashboard Cite

show abstract

“…We have used a modified version of the code developed by Gebhard et.al. [77] for this purpose. While Gebhard et.al.…”

Section: Sample Generationmentioning

confidence: 99%

“…The component black hole masses, spins, polarization angle, cosine of inclination, COA phase angle and injection SNRs are all uniformly sampled from the specified ranges. A brief explanation for each of these parameters and their significance are given in the Appendix in [77].…”

Section: Input Layermentioning

confidence: 99%

Using deep learning to localize gravitational wave sources

et al. 2019

View full text Add to dashboard Cite

Deep Learning algorithms, in particular neural networks have been steadily gaining popularity among the gravitational wave community for the last few years. The reliability and accuracy of Deep Learning approaches in gravitational wave detection, parameter estimation and glitch classification have already been proved and verified by several groups in recent years. In this paper, we report on the construction of a deep Artificial Neural Network (ANN) to localize simulated gravitational wave signals in the sky with high accuracy. We have modelled the sky as a sphere and have considered cases where the sphere is divided into 18, 50, 128, 1024, 2048 and 4096 sectors. The sky direction of the gravitational wave source is estimated by classifying the signal into one of these sectors based on it's right ascension and declination values for each of these cases. In order to do this, we have injected simulated binary black hole gravitational wave signals of component masses sampled uniformly between 30-80 M into Gaussian noise and used the whitened strain values to obtain the input features for training our ANN. We input features such as the delays in arrival times, phase differences and amplitude ratios at each of the three detectors Hanford, Livingston and Virgo, from the raw time-domain strain values as well as from analytical versions of these signals, obtained through Hilbert transformation. We show that our model is able to classify gravitational wave samples, not used in the training process, into their correct sectors with very high accuracy (> 90%) for coarse angular resolution using 18, 50 and 128 sectors. We also test our localization on test samples with injection parameters of the published LIGO binary black hole merger events GW150914, GW170818 and GW170823 for 1024, 2048 and 4096 sectors and compare the result with that from BAYESTAR and Parameter Estimation (PE). In addition, we report that the time taken by our model to localize one GW signal is around 0.018 secs on 14 Intel Xeon CPU cores. composition of these compact objects. [14][15][16][17][18][19].LIGO and Virgo began their third observation run (O3) in April, 2019 and have already detected several new GW signals. The technique used by search pipelines to identify and characterize GW signals from noise is matched filtering, which uses a bank of template waveforms of different component masses and spins [20][21][22][23][24][25][26][27]. These waveform models cover the inspiral, merger and ringdown phases of a compact binary coalescence and are generated by combining post-Newtonian theory [28][29][30][31], the effective-one-body formalism [32] and numerical relativity simulations [33]. While this approach has been very successful, the algorithms used are computationally expensive due to the fact that they probe a large parameter space with increasingly longerduration waveforms, as the low-frequency sensitivity of the interferometers improves. Also, future observation runs will involve more detectors running simultaneously over longer periods of time, th...

show abstract

“…Neural networks have been shown to achieve extremely good sensitivities in both white and real noise, comparable with matched filtering and some unmodeled methods [11,19,[21][22][23][24]. However, each study addresses quantitatively some but not all of the following points: (1) how to train, (2) how much to train, (3) robustness towards signals on which the networks were not trained, (4) differences in architectures' effects on detection efficiency/false alarm probability, and (5) applying the networks to a real search.…”

Section: Introductionmentioning

confidence: 99%

How effective is machine learning to detect long transient gravitational waves from neutron stars in a real search?

et al. 2019

View full text Add to dashboard Cite

We present a comprehensive study of the effectiveness of Convolution Neural Networks (CNNs) to detect long duration transient gravitational-wave signals lasting O(hours − days) from isolated neutron stars. We determine that CNNs are robust towards signal morphologies that differ from the training set, and they do not require many training injections/data to guarantee good detection efficiency and low false alarm probability. In fact, we only need to train one CNN on signal/noise maps in a single 150 Hz band; afterwards, the CNN can distinguish signals/noise well in any band, though with different efficiencies and false alarm probabilities due to the non-stationary noise in LIGO/Virgo. We demonstrate that we can control the false alarm probability for the CNNs by selecting the optimal threshold on the outputs of the CNN, which appears to be frequency dependent. Finally we compare the detection efficiencies of the networks to a well-established algorithm, the Generalized FrequencyHough (GFH), which maps curves in the time/frequency plane to lines in a plane that relates to the initial frequency/spindown of the source. The networks have similar sensitivities to the GFH but are orders of magnitude faster to run and can detect signals to which the GFH is blind. Using the results of our analysis, we propose strategies to apply CNNs to a real search using LIGO/Virgo data to overcome the obstacles that we would encounter, such as a finite amount of training data. We then use our networks and strategies to run a real search for a remnant of GW170817, making this the first time ever that a machine learning method has been applied to search for a gravitational wave signal from an isolated neutron star.

show abstract

Convolutional neural networks: A magic bullet for gravitational-wave detection?

Cited by 138 publications

References 71 publications

Deep-learning continuous gravitational waves

Deep-learning continuous gravitational waves

Using deep learning to localize gravitational wave sources

How effective is machine learning to detect long transient gravitational waves from neutron stars in a real search?

Contact Info

Product

Resources

About