(Original title: From physics, sociology, astronomy, medicine to chemistry, the scientific community is changing by artificial intelligence) Just as the term "neural network" enables the public to develop endless imaginations, particle physicists have begun to use artificial intelligence since the 1980s. Since almost every experiment is a pattern that finds a slight space in the countless highly similar data of complex particle detectors—which AI is good at, artificial intelligence and machine learning naturally apply to the field of particle physics. “It took us many years to convince people that it’s not magic, it’s not a juggling, it’s not a black box operation.†This was said by Boaz Klima, the first batch of Fermilabs at Fermilab in Baltimore, Illinois. One of the physicists who use this technique. In order to understand the mysteries of the universe, particle physicists need to smash subatomic particles and blow out new abnormal materials with great power. (In physics, anomalous matter refers to something that is different from ordinary matter, and it is a collective name for substances with strange characteristics. ). For example, in 2012, scientists discovered the legendary Higgs boson using the world's largest proton collider (Large Hadron Collider (LHC) in Switzerland). This fleeting particle is the key for physicists to explain how all other elementary particles get their quality. However, such abnormal substances are difficult to characterize. In the LHC, a Higgs boson appears in about one billion proton collisions, and within a billionth picosecond, it decays into other particles, such as a pair of photons or is called One-quarter particle of muon. In order to "reconstruct" a Higgs boson, physicists must discover all the more common particles to see if they can fit in with the same material from the same parent - in the typical collision process, a lot of The related particle swarm makes this task more difficult. Pushpalatha Bhat, a physicist at Fermilab, said that neural network algorithms are superior to directly filtering information from raw data. In particle detectors - this is usually a huge barrel structure made up of various sensors - photons usually produce particle sprays in subsystems called electromagnetic calorimeters. Although electrons and particles are called hadrons, there are subtle differences between their clusters and photons. Machine learning algorithms can discern differences by detecting the correlation between multiple variables of such clusters, and can also help distinguish between any pair of photons generated by Higgs after decay. "This is like a needle in a haystack," Bhat said. "So extracting the most information from the data is very important." Machine learning has not yet conquered this area. Physicists still rely mainly on the understanding of implied physics to find data related to new particles. But AI is likely to become more and more important, said Paolo Calafiura, a computer scientist at the Lawrence Berkeley National Laboratory in Berkeley, California. In 2024, researchers plan to upgrade the LHC to increase the collision rate by a factor of 10. At this point, Calafiura said, machine learning is crucial to keeping up with the data tide. 犓惴ㄈ绾 龃笾谇樾鳇 龃笾谇樾鳇/b> Social media brings billions of users and hundreds of billions of Twitter and posts to social sciences every year. The psychologist Martin Seligman realized that this also provided an unprecedented opportunity for the use of artificial intelligence to study the direction of mass communication. In the World Well-Being Project at the Center for Positive Psychology at the University of Pennsylvania, he used more than 20 psychologists, doctors, and computer scientists to screen data using machine learning and natural language processing methods to detect public and physical Health status. This is usually done by questionnaires. However, social media data is “unremarkable, cheap, and orders of magnitude larger,†Seligman said. Of course, these data first need a lot of preprocessing, but AI also provides powerful visualization tools. In a recent study, Seligman and his colleagues tracked the daily updates of 29,000 Facebook users who participated in the depression self-assessment. Using data from 28,000 of these users, machine learning algorithms found a link between the vocabulary used for the update and the level of depression. Then it can successfully predict the depression of other users based on its updated content. In another study, panelists estimated the death rate from heart attacks in counties by analyzing 1,480 tweets. Words related to anger and negative relationships are classified as dangerous causes. Data derived from social media information is closer to the true death rate than the so-called ten key factors in traditional impressions, such as smoking and diabetes. Through social media information, researchers can also predict personality, income, and political tendencies; they also study medical care, past experiences, and orientation patterns. Using Twitter data, the team even created a map of the US counties based on the happiness index, depression level, trust level, and five personality traits. "The cross-analysis of language and psychology is bound to have a revolution," said James Pennebaker, a social psychologist at the University of Texas at Austin, who says that his focus is not on content but on style. For example, scores can be predicted by observing function words used in university applications. Articles and prepositions represent dialectical thinking and higher scores; pronouns and adverbs represent narrative thinking and lower scores. According to the report, most of the 1721 drama "Double Falsehood" was written by Shakespeare. Pennebaker also found evidence that machine learning algorithms are based on factors such as cognitive complexity and rare words. Shakespeare's other works are matched. "Now we can analyze all the content that you published and written before." Pennebaker said, the result is, "More and more pictures piece together an original you." Combating Autism Genes For geneticists, autism is an annoying challenge. The genetic map shows that it has strong congenital genetic factors. However, variants of dozens of known genes that play a role in autism can only explain the cause of about 20% of cases. Finding other variants in the other 25,000 human genes and related DNA data may be helpful in explaining autism completely. So Olga Troyanskaya, a computational biologist at Princeton University, and the Simons Foundation in New York City also picked up artificial intelligence weapons. Robert Danell, the founder of the New York Genome Center and a clinical scientist at Rockefeller University, explained: “We can only do what biologists do to discover the secrets of diseases like autism. A machine can search for it. A scientist could only find 10 at the same time as the problem of success. This completely changed the rules of the game." Troyanskaya collected hundreds of data sets, including data on active genes in specific human cells, how proteins interact, and where the binding points of transcription factors and other key genome features are located. Then her team used machine learning to construct a gene interaction map and compare a few of the known high-risk genes that contribute to autism with thousands of other genes to find similarities. They published 2,500 genes that may be related to autism in last year's Nature Neuroscience. But geneticists have only recently realized that genes are not isolated. Their behavior is caused by a combination of millions of nearby non-coding genes and interacts with DNA-binding proteins and other factors. Identifying which non-coding variants may affect nearby autism genes is more difficult than finding the diseased genes themselves. Ji Zhou, a graduate student in the laboratory of Troikaka, is trying to solve this problem with AI. To train deep learning systems, the system was applied to data gathered from Encyclopedia of DNA Elements, Roadmap Epigenomics. The two projects listed how thousands of non-coding DNA affect the location of neighboring genes. The system learned which characteristics should be captured because it estimates the potential activity of unencoded DNA. After Zhou and Troyskaya published their DeepSEA study in the October 2015 issue of Nature, Xiaohui Xie, a computer scientist at the University of California, Irvine, praised this as “a milestone in applying deep learning to genome engineering. "." The Princeton team is now sequencing the impact of non-encoded genes through DeepSEA's genome for autism patients. Xie犚沧 rushes to apply AI to the genome and is more focused than autism. He hopes to classify gene mutations by studying the probability that any gene will evolve into a harmful gene. But he realized that in genomics, deep learning systems can only perform well on the data sets they train. He said: "People doubt whether such a system can reliably resolve the genome. But I think that more and more people will undergo deep learning." Access to God's Machine In April of this year, astrophysicist Michael Schawinski made a few vague galaxy pictures on Twitter and asked if any colleagues could help him distinguish the four galaxies. Colleagues said that these images look like elliptical spiral galaxies similar to the Milky Way. Some astronomers suspect that this is Schawinski's trick and ask directly, are these real galaxies simulated or modeled on a computer? In fact, neither is. Schawinski of the Swiss Polytechnic Institute in Zurich, Ce Zhang, a computer scientist, and other collaborators created these galaxies using neural networks that have no knowledge of physics. Schawinski only wanted to use this Twitter to show how realistic the neural network is. But his bigger goal is to create techniques similar to those in movies that magically make fuzzy surveillance images clear. Neural networks can make a fuzzy star map look like it was shot with a high-performance telescope, but the actual telescope may not be so good. This also allows astronomers to observe finer details. "The money used for astronomical observations amounts to tens of millions or even billions of dollars," Schawinski said. "With this technology, we can get more immediate information. †This galaxy image is generated by a generative adversarial network, a machine learning model that interacts with two opposing neural networks. One is a generator for generating images, and the other is a discriminator that tries to reduce the generation of image defects. The team of 뫌 ܔЪchawinski, who is comparable to the thorns of the thorns, has taken thousands of galaxy's real images and artificially lowered the resolution. The researchers then let the generator handle the images more intelligently in order to be able to discriminate through the discriminator. In the end, the neural network may be better than other techniques for noise reduction of galaxy pictures. Brian Nord, an astrophysicist at Fermilab, said that Schawinski’s method is machine-learning examples of a special avant-garde used in astronomy, but it is by no means the only one. At the American Astronomical Society meeting in January, Nord proposed a machine-learning method to track strong gravitational lenses: when distant galaxy's image undergoes distorted space-time during its transmission to the earth, it forms a rare light arc in the sky. These lenses can be used to measure the distance of the universe and find invisible super-concentrates. Strong gravitational lenses are visually unique and difficult to describe with simple mathematical rules. This makes traditional computers difficult to choose, but it is easy for people to grasp. Nord et al. realized that a neural network trained through thousands of shots can obtain similar perceptions. In the next few months, "In fact, there have been more than a dozen papers using machine learning to find powerful shots. Most of them rushed in." Nord said. This is only part of a growing understanding of astronomy. Artificial intelligence provides a powerful way to find and sort interesting objects in PB-level data. For Schawinski, "I think this era really becomes an era of 'Oh, God, too much data'." Neural Network Learning Chemistry Synthesis Organic chemists are all experts looking backwards. Just as chefs start with looking at the finished product and then study the specific cooking steps, many chemists start with the synthesis of the molecules they want to make and then consider how to assemble them. Marwin Segler, a graduate student at Münster University in Germany, said: "You need the right ingredients and recipes to combine them." He and others are introducing artificial intelligence into their molecular kitchens. They hope that AI can help them deal with the key challenges that the molecule generates: Select from hundreds of potential modules and connect them with thousands of chemical rules. For decades, chemists have worked hard to extract computers with pre-installed responsiveness, hoping to create a system that can quickly calculate the simplest molecular formula. However, Saigler said that chemistry is "very subtle, and it's hard for binary to cover all the rules." So Keegler, Münster's computer scientist Mike Preuss and his mentor Mark Waller all turned to AI. They programmed through in-depth neural network models instead of rigid and fast chemical reaction rules, learning the process of chemical reactions on their own from millions of examples. "The more data you provide, the better," said Segler. Over time, the model learned the best reaction to predict the steps required for synthesis. In the end, it can build its own molecules from scratch. Three people tested the machine learning program with 40 different molecules and compared the results with traditional molecular generation programs. At a meeting this year, they made relevant statements. The conventional program only used 22.5% of the 2-hour calculation window to propose a solution to synthesize the target molecule, while the AI ​​was 95%. Ceger, who is about to work for London Pharmaceuticals, hopes to use this method to improve the pharmaceutical sector. Paul Wender, an organic chemist at Stanford University, said that it is still too early to draw conclusions about the Siegler method. However, he believes that "it may have far-reaching effects," not only in the synthesis of known molecules, but in the creation of new molecules. Siegler added that organic chemists will not be replaced by artificial intelligence soon because their capabilities go far beyond predicting how the reaction will proceed. But like a GPS navigation system, artificial intelligence can lead but it cannot design and implement a complete chemical synthesis. Of course, all AI developers have to look at the six ways.