Source 丨 Pin Play(ID:pinwancool)
Author 丨 Oil vinegar
Where can you find this AI company called DeepMind?
- Go board or Nature.
Since its inception in 2010, DeepMind has always given the outside world such a "untrimmed" feeling. A classic story is that when the first three founders, Dymis Hasabis, Mustafa Suleiman and Sean Leger, founded the company, they even built a web page with the company logo, no contact number, no company address, and no polite "about us" on the website.
"In order to hire employees, founders have to rely on personal connections to convince people that they are serious and serious scientists, that they have real plans," Hassabis said in an interview.
For most people, the moment the AI company hit was not acquired by Google in 2014, but AlphaGo two years later.
Image source| The Verge
In March 2016, AlphaGo challenged then-Go world champion Lee Se-chi kudan, winning 4–1.
In May 2017, the further evolved AlphaGo played against the world's no. 9-dan Ke Jie, 3-0. After the game, Ke Jie burst into tears and bluntly said that there was no hope of winning.
The two confrontations between the top human players and AlphaGo directly pushed Go into the ERA of AI, and artificial intelligence became the sparring partner and imaginary enemy of chess players. In February this year, South Korean chess player Shin Jin-chan defeated Yang Dingxin and Ke Jie in succession, and Ke Jie complained that Shin's full chess game "has a 71% AI match rate". The latter is also blunt, revealing that he will spend 5 hours a day practicing with AI to improve his chess strength.
In the face of AlphaGo, the public felt the oppression brought about by the embodiment of artificial intelligence for the first time. This drop of ink that fell into the water, the dragonfly changed the color of the entire Go world forever.
——The Go world looks bitter and vengeful, and the "initiator" DeepMind has turned the page in advance.
After playing against Ke Jie, the Alpha Go team announced that AlphaGo will no longer participate in Go matches, look at DeepMind's progress in recent years, can not find the figure of AlphaGo, but issued a lot of papers, research directions from life sciences to nuclear fusion, and even ancient civilizations are involved.
Play counts the number of papers DeepMind has published in Nature in recent years. In the two years so far, DeepMind has published a total of 60 papers in Nature, nearly 10 more than the total number of papers in the previous three years. Of those 60 papers, 10 were in the three months of 2022 alone.
As papers become more and more dense, and interdisciplinary linkages go to heaven and earth, DeepMind looks like an idealistic academic institution that does not love to make money but is full of vitality. No wonder anyone is kidding that DeepMind is now an artificial intelligence company that lives on Nature.
Image source| "Nature"
The cover of an issue of Nature in late 2020, with the rare use of the inscription "It will change everything" to describe a research problem about proteins, was finally overcome, and it was DeepMind who conquered it.
The prediction of 3D deconstruction of proteins has plagued biologists for many years, on the one hand, it largely determines the properties and functions of proteins, and has very high research value; but because the 3D structure of proteins has hundreds of millions of folding ways, this idea has never been realized in the 50 years since it was proposed.
This puzzle, which spans more than half a century in the history of biological research, was solved by DeepMind's artificial intelligence system, AlphaFold2. The latter scored more than 92 points in the Critical Assessment of Protein Structure Prediction (CASP) system for accurately predicting protein folding structure through amino acid sequences, meaning that its predictions based on computational biology are almost as accurate as laboratory methods.
The source | Twitter
CASP Chairman John Moult bluntly said that "the great challenges in the field of computer science in the past 50 years have been greatly solved", and the academic community has not hesitated to call it "one of the most important scientific breakthroughs made by mankind in the 21st century". Google CEO Sundar Pichai, Tesla CEO Elon Musk and others are also excited about the research results.
DeepMind, on the other hand, has disappeared from the public eye. It reappeared at the beginning of this year, and it was discovered in the tokamak device that studied nuclear fusion.
The source | Wikipedia
This time in Nature, it was because DeepMind's AI achieved precise control of the magnetic field containing plasma in the tokamak.
The core condition for nuclear fusion is the need to overheat hydrogen into a plasma state at an environment of more than 100 million °C. There are two well-known ways to constrain plasma at such high temperatures — the sun with its massive mass can rely on gravity constraints, or inertia like a hydrogen bomb.
However, the objective conditions of the former cannot be reproduced on the earth, and the reaction process of the latter is uncontrollable, so the controllable nuclear fusion experiments taken by countries in the world at present mainly use the method of magnetic confinement, while the tokamak is the device for the occurrence of magnetically constrained nuclear fusion.
Due to the extremely high internal temperatures, the plasma needs to be suspended in the device by a magnetic field from 19 magnetic coils, so the control requirements for the magnetic coil are extremely high, and in an almost completely random environment, the coil may need to be adjusted thousands of times in a second. Once it comes into contact with the device during its calibration, it may lead to a weakening of the nuclear fusion reaction. Therefore, researchers did not dare to take too much risk to explore the upper limit of the core reaction of the tokamak device.
Source| Center for Science and Engineering Computing, Southern University of Science and Technology
In simple terms, what DeepMind's deep learning AI does is help researchers further precisely control the tokamak — such as accurately capturing all the variables present in a real tokamak device, and making decisions to adjust the magnetic coil in 50 millionths of a second.
"Some of the plasma shapes we're trying to force the entire unit to run the system is approaching its limits, in which case the plasma could collapse and damage the device." We wouldn't take that risk without confidence in AI," said Ambrogio Fasoli, one of the scientists at the Swiss Plasma Center, who is involved in the project, of DeepMind.
DeepMind then dropped nuclear fusion and went to ancient civilizations, this time without disappearing for long, and less than a month later, DeepMind's work was once again on the cover of Nature, with an ancient Greek stele in the background.
The DeepMind team, together with research teams from the University of Venice and Harvard University, has come up with a new way to restore and identify the age of ancient stone tablets and the content of inscriptions. Specifically, a series of codes issued by Athens during the ancient Greek period are thought to have been written before 446-445 BC, and this study accurately defines this time point to 424-423 BC, which is of great significance for historians to study the evolution of the cultural and political system of ancient Greece.
It's not uncommon for tech companies to periodically publish papers to showcase some of the research. Bart Selman, a professor of artificial intelligence at Cornell University, has said that the world's top tech giants publish at least dozens of papers at these conferences every year, and "in the field of artificial intelligence, you must present papers at mainstream artificial intelligence, machine learning, computer vision and natural language conferences with peer review."
Whether it is to show their R&D capabilities or for the purpose of talent recruitment, the paper results of technology giants such as Google and Microsoft often appear in the top meetings of artificial intelligence such as AAAI and NeurIPS or in authoritative academic journals such as Nature.
The source | Youtube
Musk has questioned the paper's practical usefulness. For the Tesla CEO, he thinks success at the academic level is relatively easy because you only need to publish a few useless papers, "and in fact, the vast majority of papers are useless."
For DeepMind, publishing papers is not only a daily action, but also always appears in a variety of fields that seem unrelated to him.
It's worth noting that the two are not unrelated, musk was one of DeepMind's early investors, and the timeline even preceded Google.
On why the AI company is becoming more and more keen on "crossover", 2019 is a node. At the time, the company was preparing to move out of the King's Cross, Google's British headquarters in London, and move on to the next phase of its growth cycle. Along with the changes, there is also the company's future development direction.
图源｜Data Center Knowledge
"The problem with reinforcement learning is that it's always on paper or just focuses on a very narrow grid world, but it's always been questioned whether these theories work when it comes to more complex scenarios in the real world."
Hassabis said the company will extend reinforcement learning to solving real-world problems in the future. The landing of the new headquarters also marks that DeepMind will turn to the basic problems surrounding organic life with the research and development strength and all the technology accumulation accumulated in the previous decade. What initially encouraged him to do so was AlphaGo's success in Go a few years ago.
In the founder's view, the collision of interdisciplinary disciplines is extremely important to the development of the scientific field, which has been implemented in the recruitment guidelines of DeepMind since its inception.
Demis Hasabis | Bloomberg Businessweek
"Glue people," Hassabis described as the talent he was eager to recruit.
"I want them to be the top talent in multiple fields at the same time, such people have the creativity to find connections between different disciplines, and when that happens, the magic happens."
In DeepMind's AI research center, in addition to mathematics, physics and neuroscience talents, biology, psychology and even philosophical backgrounds are also increasing, and the practical scientific research progress made by DeepMind in different fields in recent years is also a further enriched mapping of the disciplinary dimensions included in the team.
The source | DeepMind's official website
Going back to the study of ancient Greek steles, the synergy between humans and artificial intelligence deserves to be revisited. Will the cross-border of artificial intelligence evolve into a cross-generational plan, and even make human researchers lose their meaning?
In the course of this research, DeepMind's deep neural network Ithaca was placed in a database of Greek inscriptions provided by the Packard College of Humanities for learning. The results showed that when Ithaca was used alone to reconstruct damaged inscriptions, its recognition accuracy could reach 62%, and historians who were the reference group were only 25% accurate in the same situation.
Obviously, the introduction of deep learning tools is visible to the naked eye about the efficiency gains in the study of ancient civilizations.
But that doesn't mean AI will replace humans. Commenting on Ithaca's study of the restoration of ancient Greek steles, Charlotte Roueché, Professor Emeritus of Greek Digital Studies at King's College London, said that "this progress should not be understood as a threat to researchers in this discipline, but rather Ithaca is like a hunting dog that will look for clues for scholars".
The synergy it exhibited in Ithaca's collaboration with historians was another highly valuable finding in the study, which combined with a 72 percent accuracy rate.
DeepMind's cross-border activity is a positive signal, which shows that artificial intelligence is finding practical entry angles in more fields, and researchers in this field will be the biggest beneficiaries, and their productivity will be further liberated after getting a brand-new "steam engine".
So we have to wait for what a whole new story DeepMind — or some other AI company — will bring to Nature the next time.
bibliography
《Inside DeepMind's epic mission to solve science's trickiest problem》
Nature's Latest Cover: DeepMind AI "Next City", Tracing the Traces of Ancient Human Texts
"There is an AI to learn to control the nuclear fusion reactor, from DeepMind, on today's Nature"
Shock the Scientific Community! DeepMind AI cracks the "protein folding" problem and overcomes the great challenge of biology for 50 years
"Apple is still a "fall behind giant", and it is useless to publish papers"
This article is reproduced from Pinwancool (ID: pinwancool) with permission, if you need to reprint it for a second time, please contact the original author