In recent years, more and more artificial intelligence used in astronomical research. Depth learning needs huge amounts of data, and astronomy is the AI field to show their talents. Machines can replace human fishing in the open sea from the needle, to capture new stars, new exoplanets and even dark matter. Looking for dark matter, the machine is better than human eyes. Recently a paper "computing astrophysics and cosmology," published showed US Lorentz Berkeley National Laboratory (hereinafter referred to as "Berkeley Lab") and other agencies jointly developed by AI deep learning framework, able to explore the universe of dark matter sign.
Identify the "gravitational lens", AI meritorious
Looking for "gravitational lens" is the basic method for the distribution of dark matter research. The great mass of the object as a lens would distort light passing identify such distortions can be captured without emitting the mass thereof.
Paper display, Berkeley Lab established deep learning AI framework CosmoGAN, gravitational lensing can be analyzed associated with dark matter. It can create a high fidelity, weak gravitational lens converges FIG.
Once upon a time, looking for "Gravitational" desired analog and data processing is troublesome. 20 scientists spent several months can only see a small piece of space images. Physical modeling requires billions of hours of calculation, taking the number of megabytes of disk space.
Advances in neural network provides an opportunity. Berkeley Lab team led by the introduction of a "generative against the network (GANs)". The researchers Mustafa said: "there are other ways to get to learn the depth chart from the convergence of many images, but compared to competing methods, GANs generate very high resolution image, while still highly efficient neural network."
Now, astronomers can analyze much larger CosmoGAN sky, faster.
CosmoGAN not the only progress in astronomy deep learning neural network. For example, the University of Toronto using deep learning technical analysis of satellite images of the moon's craters, P8 supercomputer neural network discovery 6000 new crater within just a few hours, over the past few decades the number of craters found in humans twice. University of Illinois at Urbana - Champaign use depth study and analysis to detect gravitational waves black hole collisions. AI everywhere in astronomy.
Too much data, not the machine could not handle
Over the past few years, most of all in the direction of astronomy attempt to use artificial intelligence. Considering the astronomical data to be processed as much as it is a natural idea. Let the machine to analyze clues practice, not so, the future of astronomy will not operate.
Recently held 2019 GPU Technology Conference attracted AI researchers all over the world. The General Assembly invited the University of California, Santa Cruz astronomer Brent Robertson speech, he said: "Astronomy is a new outlet data revolution." Robertson believes that a new generation of astronomical instruments must meet the new generation of software driven by the depth of learning.
For example, the telescope is expected to tour throughout the day (LSST) in the large-diameter run three years later. It patrol south half of the universe 37 billion galaxies, it generates a long decade of uninterrupted video. LSST is equipped with a 3.2 billion-pixel camera, night produced 25TB of data, all data is now equivalent to the contribution of advanced telescopes life.
Another example square radio telescope array (SKA). It is all over the world, part of the antenna deployed in eight countries in southern Africa, there are more than one million antennas in Australia and New Zealand. Its original data reaches 5,000 pieces per day PB, after treatment there are about 50 PB.
"The Dark Energy Survey" star map prepared hundreds of millions of galaxies; "Gaia" satellite mapping Milky Way billions of stars; "Zwicky" project per hour can scan 3750 square degrees of sky. In China, FAST amount of data each day to reach 150TB; Guoshoujing telescopes spectrum of 9.01 million, is the world's largest astronomical spectra library ......
Capture mode humans can not see
More and more data, scientists are trying to polymerization. But the GPU conference, Robertson said that the future produce large amounts of data from several large telescopes together, after polymerization complex to humans can not directly use. The University of California, Santa Cruz scientists trying to solve this problem. Morpheus depth learning framework to create a doctoral student Department of Computer Science, the telescope can be based on the raw data, pixel by pixel classification objects.
To better study the formation of galaxies UCSC scientists also used AI. In a study they published in early 2019, scientists used computer simulations of galaxy computer training, learning it three key stages of galaxy evolution. Later, computer analysis of galaxy images from the Hubble Space Telescope, showing surprisingly good.
Artificial intelligence applied to face recognition, after the massive data training, according to a photo, recognize this person make-up and old time look. The universe many images can also be used the same method to classify.
"Deep learning can seek mode, the machine can see a very complicated mode, and humans can not see." Participated in the study library scientist David said, "We hope to further test this approach. In the proof of concept study, the machine seems We succeeded in finding different stages of the simulation to determine the evolution of galaxies in the data. "
Help astronomers find another solar system
A report by the end of 2018 shows that Google artificial intelligence force, found a new planet Kepler exoplanet observations from the database. The planets are hard to find. Kepler space located observed 145,000 sun-like stars, small changes in the brightness of stars to detect planets. Record four years of data, including approximately 35,000 suspected planets record. Astronomers using the machine to identify the combination of the human eye, but the darkest weakest signal is often overlooked.
With the help of Google AI, we discovered Kepler Kepler 80g and 90i two new planets. Kepler also to be recognized as the first 90 have at least eight planets outside the galaxy.
Neural networks and machine learning processed 14 billion data points, after successfully screened out candidates.
NASA and Google said that the future of new technologies will find more exoplanets. NASA astronomers said do not worry about unemployment. NASA scientists explained Jesse Patterson Road, said before the data is provided to the neural network, astronomers need to be classified, so that artificial intelligence can learn to analyze the new information.
Road Robertson said: "In the future AI will definitely work together and astronomer, has become an indispensable tool."
Of course, machine learning has also brought a "black box" Risk: We got the answer, but we do not know why the machine so determine, perhaps the answer is wrong. Machines make mistakes. Astronomers will continue to train and adapt to it.
Further reading
Expert Reviews
Deep learning does not yet have a "physical intuition"
Indeed, artificial intelligence has now deep into the various branches of astronomy astrophysics. After Currently, the US Lorentz Berkeley National Laboratory, the use of deep learning, can quickly be distributed according to the three-dimensional density of the universe, to determine fundamental constants dark matter and dark energy and other cosmological, they found that the application of artificial intelligence, statistics error than the previous use of traditional statistical approaches small lot. In addition, we also use depth study of the early universe look for hydrogen, carbon at low signal to noise ratio of the spectrum, it was found also easy to use than traditional methods.
Meanwhile, astronomers also apply deep learning, help us to determine three-dimensional positions of celestial bodies, the distance, and then outlines the large scale structure of three-dimensional space. It was found that the depth of learning in terms of data mining information, probably stronger than traditional methods Before we used. Artificial intelligence has also been applied to the field of Google's exoplanet detection, and successfully detected several exoplanets ...... can say that artificial intelligence is now being widely used in cutting-edge astrophysics.
But from a physicist's point of view, perhaps based on artificial intelligence has its limitations depth learning. This limitation is that it can only be based on data, to play a role in the already very clearly defined specific areas. Only under the guidance of physicists, the statistic error bars made smaller, more precise estimate a certain amount, but now we can not guide us to discover new laws of physics behind the data. Not only with human beings based on beauty, symmetry and simplicity "physical intuition."
Let me cite one simple example, such as Kepler based on observations of Tycho, Kepler's third law can be found, and now even the best machine learning, artificial intelligence algorithms may also be difficult based on the same data, repeat this discovery .
So I think the essence of deep learning applications in astronomy, is still limited and do better statistical fit this regard.
Dalian gynecological examination fee http://mobile.403yy.com/