Teaching the machines to look for fast radio bursts
Teaching the machines to look for fast radio bursts
Title: Fast Radio Burst 121102 Pulse Detection and Periodicity: A Machine Learning Approach
Authors: Yunfan Gerry Zhang, Vishal Gajjar, Griffin Foster, Andrew Siemion, James Cordes, Casey Law, Yu Wang.
Institution of the first author: Dept. of Astronomy, University of California Berkeley
State: Submitted to ApJ, open access.
Astrobite original: Teaching Machines to find Fast Radio Bursts by Joshya Kerrigan.
Today's astrobite combines two fascinating and independent themes due to a very important result: automatic learning and rapid radio bursts (FRBs). The field of machine learning is moving at an unprecedented pace with new fascinating results. The FRBs have a completely unknown origin and are preparing experiments to detect more of them. So let's get to the point and let's look at how the authors of today's article managed to have a machine identify Rapid radio bursts.
Convolutional neurological networks
Let's start by presenting the technique and machinery that the authors used to find these signals. The field of automatic learning (machine learning in English, N. del T.) is exceptionally glowing at this time, and with new knowledge being introduced almost daily into the best machine learning algorithms, dissemination to nearby fields is accelerating. This is not, of course, an exception for astronomy (radio or other), where data sets grow to be extraordinarily large and intractable for classical algorithms. I present to you convolutional neuronal network (CNN, for its acronym in English, N. del T.), the algorithm of automatic learning most frequently used for the understanding and prediction of data with spatial characteristics (ie images).
And how does one of these algorithms work? Well, a basic starting point would be the traditional neural network, but I'll leave that Explanation to someone else. A neural network can take as few as many input parameters, which do not necessarily have to be spatially related to each other, the opposite of a CNN, which is very suitable for images. (Note: you can also have CNNs 1d or 3d). These images, of course, have characteristics, those that when combined are important to identify what is in the image. Take for example Figure 1: a dog has features such as drooping ears, or a large mouth with a protruding tongue. A CNN learns part of all these characteristics from a set of training data that is provided, with a settled truth. In the case of figure 1, the prediction can be a dog, a cat, a lion or a bird. These characteristics are learned by varying the spatial scales as the images provided are successively convolved and the prediction is compared to its known description with any correction propagated backwards to update those characteristics. This last step is the part of the training, which as you may notice, is the same process as in a non-convolutional neural network. Then, armed with this brand new and fast classifier, we can move towards an understanding of what we will be predicting.
Figure 1. An example of a convolutional neuronal network. An input image is sequentially convolved through several convolutional layers, where each successive layer learns unique traits, which after training, are finally used to make a prediction based on a set of labels. Adapted from: https://www.kdnuggets.com/2018/06/topological-data-analysis-convolutional-neural-networks.html.
Rapid radio bursts
Figure 2 Simulated FRB pulses in the GBT time-frequency data. The pulses are simulated with a variety of parameters in order to make the CNN as robust as possible. Figure 1 of the article.
We have covered the FRBs in other astrobites in the past (one,two Y 3), and with each new publication, we seem to get closer and closer to finding the source of these mysterious radio signals. A quick introduction to FRBs is that they are bright bursts in radius that last milliseconds, seen in time-frequency data from radio telescopes. These bursts have unique characteristics that distinguish them from other radio signals and that will be important in understanding how the authors developed a set of training data for the predictions in the article. These characteristics consist of a dispersion measure, time of arrival (DM and TOA, respectively by its acronym in English), width and pulse width (there are others, but I will highlight these as the most important features). The DM is one of the most interesting features of an FRB, since it is the one that indicates that the FRBs are cosmological. DM is measured from the dispersion of the signal in time and frequency as it travels through an ionized medium, which in this case is the intergalactic medium. This is the curve seen in Figure 2, which delays the signal to later times while moving to lower frequencies. TOA is the moment when the signal arrived in the observations, amplitude is the flux density of the signal and pulse width is the width at 10% of the maximum amplitude.
By using all these features to define a set of training data, the authors simulated many different types of FRBs, all with their own unique values. This is important because having a large and robust training data set means that the neural network is more likely to be able to make robust predictions.
Putting CNN to work
Now we have all the components: a convolutional neuronal network, a robust training data set and a monumental amount of data from the Green Bank telescope (GBT for its acronym in English). The authors seek to probe the file data of the now well-known FRB 121102, which has had a history of being a repetitive FRB. This means that FRB 121102 is a surprising resource for understanding FRBs because we can take many measurements.
Figure 3 Distribution of the different parameters of the FRB 121102 pulses discovered in the GBT data file. Understanding how these parameters relate to each other can give us clues to the nature of FRB 121102. Figure 4 of the article.
Using many hours of GBT file data, the authors put CNN to work to predict if there would be additional FRB 121102 FRB pulses that might not have been noticed because the signal was weak or simply unnoticed. between the large amounts of data. Successfully they found 72 additional pulses from FRB 121102! And if it was not interesting enough, more than half of these newly discovered pulses occurred within the first half hour of the data set. This gives a total count, including the oldest signals, of 93 pusls FRB.
The detection and additional measurement of these pulses is certainly important. As we have said in other previous astrobites, the origin of these bursts is almost completely speculative and we need to get as many measurements as we can to discard or restrict the potential cosmological sources. Having a repetitive FRB with which to start collecting data, such as the distributions in Figure 3, is great for understanding the environment of the FRBs that should affect these parameters. Being optimistic, with the continued development of these CNNs and other machine learning techniques, we will see an explosion of FRB detections.
!function(f,b,e,v,n,t,s)
{if(f.fbq)return;n=f.fbq=function(){n.callMethod?
n.callMethod.apply(n,arguments):n.queue.push(arguments)};
if(!f._fbq)f._fbq=n;n.push=n;n.loaded=!0;n.version='2.0';
n.queue=[];t=b.createElement(e);t.async=!0;
t.src=v;s=b.getElementsByTagName(e)[0];
s.parentNode.insertBefore(t,s)}(window, document,'script',
'https://connect.facebook.net/en_US/fbevents.js');
fbq('init', '369524843414444');
fbq('track', 'PageView');
(function(d, s, id) {
var js, fjs = d.getElementsByTagName(s)[0];
if (d.getElementById(id)) return;
js = d.createElement(s); js.id = id;
js.src = "http://connect.facebook.net/es_LA/sdk.js#xfbml=1&version=v2.6";
fjs.parentNode.insertBefore(js, fjs);
}(document, 'script', 'facebook-jssdk')); .
SOURCE LINK ERESVIRAL.COM https://www.beviral.online
Comentarios
Publicar un comentario