Generalised gravitational wave burst generation with generative adversarial networks

Abstract

We introduce the use of conditional generative adversarial networks (CGANs) for generalised gravitational wave (GW) burst generation in the time domain. Generative adversarial networks are generative machine learning models that produce new data based on the features of the training data set. We condition the network on five classes of time-series signals that are often used to characterise GW burst searches: sine-Gaussian, ringdown, white noise burst, Gaussian pulse and binary black hole merger. We show that the model can replicate the features of these standard signal classes and, in addition, produce generalised burst signals through interpolation and class mixing. We also present an example application where a convolutional neural network (CNN) classifier is trained on burst signals generated by our CGAN. We show that a CNN classifier trained only on the standard five signal classes has a poorer detection efficiency than a CNN classifier trained on a population of generalised burst signals drawn from the combined signal class space.

Simulating Transient Noise Bursts in LIGO with Generative Adversarial Networks

“…GAN can learn the underlying distribution of a population to produce artificial examples from Gaussian noise. With this idea in mind, the authors in [36] employed a conditional GAN to burst signals, allowing them to generate multiple classes of signals with the same algorithm and to interpolate through different classes, creating mixed signals. The powerful generation capability of GAN leads the authors to foresee that it could be applied to generate artificial glitches.…”

“…The structure of the glitch is fully recovered and allows to reveal the detection capability of ALBUS. As suggested in [36], when learning different classes of glitches, we could also interpolate between them to generate hybrid classes. This hybrid dataset could be employed to discover unknown classes of glitches and improve the efficiency of detection algorithms.…”

Generative Adversarial Networks Applied to Synthetic Financial Scenarios Generation

“…Introduced in 2014, by Ian Goodfellow et al [8] to circumvent the intractability of probabilistic computations encountered by Deep Generative models during maximum likelihood estimation, Generative Adversarial Networks (GANs) have shown groundbreaking results in image generation. Since their creation they have been adapted to a broader range of applications from speech synthesis (Bińkowski et al [2]) to gravitational burst generation (McGinn et al [14]) or even dynamic environments simulation (Kim et al [12]).…”

Source-agnostic gravitational-wave detection with recurrent autoencoders

“…On the other hand, GW detection comes with the need of going beyond signal signatures that can be emulated. McGinn et al [25] discusses how to generate 'unmodeled' waveforms, which could then be used to train a supervised algorithm without the use of templates.…”