For the first time, researchers used neural networks to analyze gravitational lenses, analyzing key distortions in time and space, and their analysis speed was 10 million times faster than traditional methods. According to foreign media reports, the researchers use a similar brain “neural network” analysis of time and space in the key distortions, the analysis speed than the traditional method of 10 million times faster.
This latest study uses an artificial intelligence system to detect the gravitational lens phenomenon in the Hubble Space Telescope image and to simulate the image. This process allows the researcher to better observe the quality of the galaxy Distributed, and offers “close-up” of distant galaxy celestial bodies. (Laurence Perreault Levasseur), a researcher at the Kafley Institute for Astrophysics and Cosmology (KIPAC), said: “Usually, the number of work needs to be analyzed Weeks to complete, which requires expert technical support and computing needs, but in a fully automated way, the ‘neural network’ can only be completed in a fraction of a second. “
Accidental arrangement of dense objects And galactic backgrounds can produce gravitational lenses – when the light around the universe in the background of the quality of the formation of a background when the natural magnification, resulting in light ring distortion, sometimes called “Einstein ring”, you can analyze through the analysis of distant objects The system itself and passing the mass of its front celestial bodies, which is very convenient for understanding dark matter, although scientists can not directly observe the dark matter, but it can be used as a focus on the study of the background of the galaxy “lens”.
, Analysis of the system to master the properties of celestial bodies is still a long and tedious process.
Researchers say that analyzing a single gravitational lens takes several weeks or months, while the “neural network” finds lens properties in just a few minutes.
The working principle of neural networks is to combine artificial intelligence systems with the brain’s special inspiration-inspired structure, combined with millions of billions of instance attributes, so help researchers understand how to identify other conditions Attributes. For example, the “neural network” system can accurately identify a large number of dogs in a photo without having to ask the researchers what details the neural network needs to be aware of.
Hubble Space Telescope Shoots the galaxy in a process called a gravitational lens “Bend” around the dense background objects, the researchers use these images to test the performance of the neural network to understand the characteristics of the gravitational lens.
At the same time, “neural network” can also be used to complete more complex tasks, such as: Google developed the “Alpha Dog (AlphaGo)” program, the program can run a large number of procedures and analysis, and even beat the world champion. In contrast, traditional computer programs do not have a perfect calculation of Go because there are many possibilities for Go. In this study, the researchers found that the latest design of the neural network system in a day can be analyzed and processed about 500,000 simulated gravitational lens images, and then they test the new lens on the network system, found to be able to Very fast and accurate analysis. Yashar Hezaveh, a researcher at KIPAC, says the neural network we tested includes three open neural networks and one of our newly developed neural networks that are capable of measuring The properties of each gravitational lens, including: the quality distribution status, and the degree of magnification of the background galaxy image.
Researchers point out that although neural networks are used in the field of astrophysics before, but are rarely used in complex areas of the field. For example, they are used to identify whether the image contains a gravitational lens, but not for analyzing it. “It’s like our newly developed neural network not only captures the image of the dog from a large number of photos, but also returns information about the weight, height and age of the dog.”
Although these studies are Through the high-performance computing cluster to complete, but the researchers said that with less calculation can be completed, such as: in the laptop, and even mobile phones can also be processed to complete. As more and more astronomical data need to be analyzed and analyzed, computing systems like this will be a very critical tool. Roger Blandford, a researcher at KIPAC, points out that neural networks have been used to solve astronomical problems and that the results are mixed. But the latest algorithm combined with modern image processing units (GPUs) can create very fast and reliable results, just as the research report demonstrates the resolution of gravitational lens problems. It is highly optimistic that this will be the choice of more data processing and analysis issues in astrophysics and other fields. At present, the latest research report published in the August 30 issue of “Nature” magazine. (Allure)