Gravitational lenses are the object of long-term search and research in astronomy, but scientists have been unable to help. With deep learning and computer vision technology to process the large amounts of data generated by telescopes, scientists will be able to expand their understanding of the universe through numerous gravitational lenses.
If Earth is in the same line with two other galaxies, when a galaxy or galaxy blocks another galaxy behind it, the gravity of the first galaxy will bend the light from the second galaxy. Gravity lenses will appear. The gravitational lens effect makes the first galaxy a magnifier that observes the second galaxy on Earth. However, the recognition of gravitational lenses has proved to be a major challenge.
By accurately recognizing the gravitational lens and then analyzing the telescope data, scientists not only can better observe the more distant galaxies, but also can essentially understand the unknown matter form, dark matter, which may be spread throughout the universe.
Says Yashar Hezaveh, a postdoctoral fellow at the NASA Hubble Telescope Project at the Stanford University's Kavli Institute for Particle Astrophysics and Cosmology: “With gravitational lenses we can learn a lot about scientific knowledge. We can use these data to study the distribution of dark matter and the formation of stars and galaxies. ."
In-depth exploration of deep learning
Not too long ago, scientists still analyzed images through a lot of complicated computer code. This method requires a large number of supercluster-related calculations, and also requires a lot of man-made operations. However, everything changed when Hezaveh and his research team decided to use computer vision technology and neural networks.
Laurence Perreault Levasseur, a postdoctoral researcher at Stanford University, and coauthor of the topic "Fast Automated Analysis of Strong Gravitational Lenses with Convolutional Neural Networks" (download link: https://arxiv.org/abs/1708.08842), said: "We I didn't think about how effective it would be, or whether it would work."
We can also think of a gravitational lens as a magic mirror. The challenge is to eliminate the effect of mirror distortion and find the true image of the object in front of the mirror. The traditional approach is to compare the observations with a large dataset of simulated images of the same object seen in different mirrors to find more similar results to the data.
The neural network can directly process the image and find the real image without having to compare it with a large number of analog images. In principle, this method can speed up the calculation. However, training deep learning models to understand how fluctuations affect material activity and our observations also requires strong computing power.
After Hezaveh and his team used GPUs to analyze data, they were able to quickly and accurately reveal new insights about the universe. Using Stanford's Sherlock high performance computing cluster (based on NVIDIA GPUs), the team trained models up to 100 times faster than when using CPUs.
This will gain a deeper understanding of the gravitational lens and provide ample material for those who desire a deeper understanding of the universe.
Perreault Levasseur pointed out: "Using this tool can answer many scientific questions."
Fully "together" gravitational lens
Of course, to analyze the gravitational lens data, we must first find the gravitational lens, which is exactly what the scientists of the three European universities are trying to solve.
As part of the Kilo-Degree Survey (KiDS), an astronomical observation project aimed at a better understanding of the dark matter and mass distribution in the universe, the University of Groningen, the University of Naples and the University of Bonn Researchers at the University of Bonn have been using deep learning methods to identify new gravitational lenses.
Carlo Enrico Petrillo is a related deep learning paper "Finding strong gravitational lenses in the Kilo Degree Survey with Convolutional Neural Networks" (download link: https://academic.oup.com/mnras/article-abstract/472/1/1129/4082220 The co-author of the book, he pointed out that even if this observational project only observes a small slice (about 4%) of space, it still found up to 2,500 gravitational lenses when implementing KiDS using AI technology.
However, before we achieved this achievement, we faced a major challenge, namely the lack of a vital training data set that is usually required for deep learning applications. Petrillo said that the team’s strategy is to simulate the arcs of light and rings around the gravitational lens and then add them to the image of the real galaxy.
"In this way, we can use the image of the observations to get all the specific characteristics (such as resolution, wavelength and noise) to simulate the gravitational lens." Petrillo said.
In other words, the team considers this issue to be one of the binary classifications: galaxies surrounded by light arcs and rings that match the simulation results are marked as lenses, and unmatched galaxies are labeled as non-lenses. As the network continues to learn from each simulation, researchers can narrow the range of candidates. The team's published paper points out that this method allows them to reduce the 761 candidates to a list of 56 suspected gravity lenses from the beginning.
NVIDIA GPUs help to achieve this result by significantly reducing the time required to compare images to analog results. It takes 25 seconds to compare a batch of images on the CPU, and the GPU increases the speed by 50 times.
"Using the CPU will make my job miserable," he said.
The flood of data hits
With the continuous innovation of telescopes and deep learning technologies, the volume of data for gravitational lenses is expected to increase significantly. For example, Petrillo pointed out that the European Space Agency’s Euclidean telescope is expected to generate dozens of PBs of data while Chile’s large-calibre all-day inspection telescope will produce 30 TB per night.
This means that it will need to process large amounts of data, find many gravitational lenses and recognize new space boundaries. This will also pose new challenges for scientists.
Petrillo said: "Finding a large number of gravitational lenses means that the formation and evolution of galaxies can be accurately presented, and the nature of dark matter and the structure of the space-time continuum itself can be deeply understood. We need to use an efficient and fast algorithm to analyze all these data, and machine learning is undoubtedly It will be a common concern for astronomers."
Digital Signage for office
digital signage for office
Jumei Video(Shenzhen)Co.,Ltd , https://www.jmsxdisplay.com