This year, we continue to reap the unexpected joy brought by AI
In the past few years, when it comes to artificial intelligence, people’s first reaction is global popularity, Internet hotspots, and foreseeing the future. But now, artificial intelligence has long faded away from mystery. Voice recognition, image recognition, intelligent image reading, virus sequencing, drug design... are all Has become a handy application.
In particular, it is worth mentioning that virtual reality and autonomous driving will usher in a blowout in 2021.
Under the strong wind of "Meta Universe" this year’s hottest technology concept, virtual reality technology has ushered in a new turning point in the development of the industry. Investment and financing confidence and activity in related fields have entered a new round of climax. The number of investment projects and total capital The volume has risen sharply, the global VR/AR head-mounted display equipment shipments have grown rapidly, and the accumulation of industrial factors such as policies, funds, and talents has accelerated.
This year is equally significant for autonomous driving.
Apple, Xiaomi, Huawei, Didi, etc. announced "building cars"; Baidu and Xiaoma Zhixing became the first batch of companies approved to carry out commercial pilot services; the country's first autonomous driving service commercialization pilot was implemented in Beijing, and the industry is in vector Production and absolute safety launched an impact.
Looking back at this year, towards the frontier and practical directions, artificial intelligence and its practitioners are running wildly.
Tesla crash attracts attention
The safety issue of autonomous driving has been brought to the fore
On March 17, a domestic Tesla Model 3 turned for no reason under the autopilot assisted state, the vehicle crashed and the front of the car was almost scrapped, but none of the eight airbags in the car opened. The Tesla technical director replied that there was no The impact hits the trigger point, so the airbag does not pop out, and the vehicle has no problem.
This is not the first time that Tesla has had a similar accident.
In 2019, foreign media reported a Tesla accident. According to the victim's lawyer, when the owner's Model 3 hit the guardrail, the airbag did not open, and the owner claimed that Tesla did not cooperate with the investigation.
Tesla's accident once again brought the safety of autonomous driving to the forefront.
In fact, 2021 is the year of domestic autonomous driving. Huawei entered the market to build cars. Baidu and Xiaoma Zhixing became the first batch of companies approved to carry out commercial pilot services. The country's first commercial pilot of autonomous driving services was implemented in Beijing. Domestic autonomous driving has moved from a test demonstration to a commercial pilot, and autonomous driving has officially entered the "second half".
At the same time, the infrastructure for autonomous driving has been basically completed. Various localities have actively promoted supporting measures such as computing centers, 5G networks, edge computing, vehicle-road collaboration, and high-precision geographic data. Various L2-L4 autonomous vehicles have begun to walk out of the closed road test test site. , Embarked on a real city road.
As the number one issue of autonomous driving, safety is worthy of caution, and it is also a key element that affects the prospects of industry companies.
Autonomous Agents and Human Debate
AI begins to have the ability to participate in complex human activities
Artificial intelligence has made another success in the field of human expertise, and it can be debated with humans.
The British "Nature" magazine published an update on artificial intelligence on March 18: Scientists reported an autonomous agent that can engage in competitive debates with humans. This "debtor project" system can engage in live debates with people. The system can scan an archive of 400 million news reports and Wikipedia pages, then organize the opening remarks on its own and refute the arguments on its own.
This is considered to be fundamentally different from the previous challenge of artificial intelligence to humans.
Although in the end the human debater was judged to win, this demonstration shows that artificial intelligence has begun to have the ability to participate in complex human activities.
This can't help but make people wonder, where will the next step of artificial intelligence go?
The world's fastest AI supercomputer starts
Stitching the largest 3D map of the universe ever
On May 27th, Perlmutter, known as the world's fastest artificial intelligence workload supercomputer, was announced.
This supercomputer has 6144 Nvidia A100 tensor core graphics processors and will be responsible for stitching the largest visible universe 3D map ever, and it is expected to reveal the secrets of dark energy.
In physical cosmology, dark energy is an imperceptible form of energy that fills space and increases the expansion speed of the universe.
The dark energy hypothesis is one of the most popular interpretations of the observation results of the accelerated expansion of the universe.
Nvidia senior product marketing manager Dion Harris said that in terms of 16-bit and 32-bit mixed-precision mathematical operations used by AI, the Perlmutter supercomputer is currently the world's fastest system.
Mankind has worked hard to learn more about the unknown side of the universe. With the "extraordinary tool" of AI, this effort may be able to bear fruit faster.
Enlightenment 2.0 released
China's trillion-parameter model breaks multiple records
At the opening ceremony of the 2021 Beijing Zhiyuan Conference held on June 1, Enlightenment 2.0 was released.
It has shown explosive growth in model scale, reaching 1.75 trillion parameters, setting a record for the world's largest pre-training model.
As a language with a large number of people in the world, Chinese has not been a super-large-scale pre-training model based on it before.
In March, China's first super-large-scale pre-training model, Enlightenment was born, and the Chinese pre-training model entered the "Large Model" lineup.
The release of Enlightenment 2.0 marked the refreshing of a number of related records.
At present, the training of language models has moved from "large model refining" to "large model refining" stage. Massive models have become the focus of attention in the industry.
From the 175 billion parameter GPT-3 to the trillion-level Switch Transformer, the parameter record is constantly being refreshed, and the scale of the language model is getting larger and larger, as if there is no end.
The big fire GPT-3 can compose poems, chat, and generate codes. The parameter scale reaches hundreds of billions, which is close to the number of human neurons.
In October, Microsoft and Nvidia jointly released the Megatron-Turing Natural Language Generation Model (MT-NLG), which has 530 billion parameters and claims to have won the two titles of "biggest" and "strongest" in the monolithic Transformer language model world at the same time.
Exploring large models is a continuous process. Scientists hope that larger and larger models can lead to the holy grail of AI-general artificial intelligence.
Alphafold2 predicts protein structure
Bring a revolutionary impact in the field of life sciences
On July 16, the British "Nature" magazine published a new study in structural biology. The artificial intelligence company DeepMind's neural network Alphafold2 predicted the protein structure to achieve atomic level accuracy.
The problem of protein folding is considered to be one of the important scientific frontier issues that humans need to solve in the 21st century.
The study of protein structure helps to understand the role of protein, understand how protein performs its biological function, and understand the interaction between protein and non-protein, which is very important for biology, medicine and pharmacy.
For more than 50 years, researchers have been trying to predict the three-dimensional structure of a protein according to its amino acid sequence.
However, the accuracy of the currently used calculation methods is limited, and the experimental methods require very high manpower and time.
In fact, in the past half a century, humans have analyzed the structure of more than 50,000 human-derived proteins. About 17% of the amino acids in the human proteome have structural information, and the structure predicted by AlphaFold2 has increased this number from 17% to 58%, because the proportion of amino acids without a fixed structure is so large, 58% of the structure predictions are already close to the limit.
At the end of this year, my country’s self-developed deep learning protein folding prediction platform TRFold has good news. Based on the protein test set of the 14th International Protein Structure Prediction Competition in 2020, its score is second only to AlphaFold2, ranking second in the world. This is currently all public in China. With the best results in protein structure prediction models, my country's performance in the field of computational biology ranks first in the world.
Shi Yigong, a biophysicist and president of West Lake University, once spoke highly of the performance of Alphafold2: This is the biggest contribution of artificial intelligence to the field of science, and it is also one of the most important scientific breakthroughs made by mankind in the 21st century.
There is no doubt that the research of artificial intelligence predicting protein structure has and will continue to cause revolutionary influence in various branches of life sciences, and it will gradually appear in the next few to ten years.
Meta Universe attracts global companies to compete
The key to artificial intelligence technology or Chengyuan Universe landing
Which technology concept is the hottest in 2021?
The answer is basically non-controversial: meta universe.
This term from science fiction novels in the 1990s has become the strongest concept that giants have been competing for, capital staking, and the streets have been hotly discussed this year.
At present, there is no universally accepted definition of the meta universe, which also creates sufficient extensibility, inclusiveness and interpretability for it.
The meta universe is like a basket, which packs all digital technologies such as augmented reality, cloud computing, digital twins, artificial intelligence, and blockchain. Among them, artificial intelligence and meta universe are inextricably related, such as deep learning, computer vision, The mature application of artificial intelligence technologies such as natural language processing is the key to the landing of the meta universe.
Although it is still at the conceptual stage, the huge commercial potential that may exist in Metaverse has attracted major companies to bet.
Nvidia officially launched Omniverse, a virtual work platform against Meta Universe, Japanese social platform GREE launched Meta Universe business, Microsoft worked hard to create an "enterprise meta Universe", Huawei entered the Meta Universe track from AR, Tencent, Baidu, ByteDance, NetEase, etc. All entered the meta universe from the path they are good at, and the three major operators have also fully invested in the meta universe layout.
In July, DeepMind created a meta-universe for AI-XLand.
After 5 generations of training, the intelligent body can play about 700,000 independent games in XLand's 4000 independent worlds, involving 3.4 million independent tasks. The last generation of the intelligent body has gone through 200 billion training steps.
Synthetic neural signals give AI "thinking"
Brain-computer connection to AI devices may usher in new developments
Once AI has "thinking", will it in turn manipulate human behavior?
Thinking about this problem now is no longer unfounded, because the AI "fake thinking" has already happened.
GAN refers to a generative confrontation network, which is a deep learning model and one of the unsupervised learning methods used by artificial intelligence in the process of big data learning in recent years.
On November 18, a paper published by Wen Shixian, a Chinese Ph.D. team at the University of Southern California, in the journal Nature, showed that researchers used brain-computer connection devices to conduct brain-computer interface training on two monkeys as test subjects.
They let the two monkeys in the experiment play a snake-eating game and a treadmill, and then collect the motor control neural signals from them, and then synthesize a large amount of neural activity data through the generator and discriminator in the GAN for the next step. test.
Through contact or implanted devices, GAN only needs to collect the motor control nerve signals emitted by the monkeys in the experiment, and can automatically generate similar nerve signals that may manipulate behavior in various other situations, and then teach these to the AI In this way, AI has its own "thinking".
Researchers have found that this technology shortens the training time for brain-computer interface systems to extract and analyze brain signals by a full 20 times.
They also mentioned in the paper that although this study only collected neural signals from monkeys, this model should also be applicable to the analog generation of human neural signals.
Researchers believe that this "synthetic thinking" approach can have a wider range of uses, especially in brain-computer connection with AI devices.
But if AI can "fake thinking", what this brain-computer connection will bring to future humans is obviously still inconclusive.
"Artificial Intelligence Ethical Issue Recommendation" released
The first global regulatory framework focuses on the healthy development of AI
Artificial intelligence governance is a problem that accompanies the development of artificial intelligence.
On November 25, local time, UNESCO officially launched the "Artificial Intelligence Ethical Issue Recommendation", which was collectively adopted by UNESCO's member states and is the first global normative framework on the subject of artificial intelligence.
The proposal aims to promote artificial intelligence to serve humans, society, the environment, and ecosystems, and to prevent its potential risks.
The proposal contains the principles that should be followed to regulate the development of artificial intelligence and the areas of artificial intelligence applications under the guidance of the principles.
According to UNESCO, the proposal has 29 pages and defines the necessary basic projects to guide the construction of artificial intelligence to ensure the common values and basic principles of the healthy development of artificial intelligence.
The proposal calls for more action beyond the measures that technology companies and governments have taken to ensure that the public is more protected by ensuring the transparency of the use of AI, the ability to act, and the protection of personal data.
The proposal also promotes ensuring that artificial intelligence becomes a more important tool for tackling climate change and solving environmental problems.
The proposal requires the government to assess the direct and indirect impacts of artificial intelligence systems on the environment, including its carbon footprint, energy consumption, and the environmental impact of raw material extraction.
AI discovers two new conjectures in mathematics
Artificial intelligence expands the scope of application in frontier fields
The field of artificial intelligence is getting bigger and bigger, this time it is mathematics.
The British "Nature" magazine published on December 1 a machine learning framework developed by the artificial intelligence company DeepMind, which has helped discover two new conjectures in the field of pure mathematics.
This research shows that machine learning can support mathematical research. It is also the first time that computer scientists and mathematicians have used artificial intelligence to help prove or propose complex theorems in mathematical fields such as knot theory and representation theory.
One of the key goals of pure mathematics research is to discover the laws between mathematical objects and use these connections to form conjectures.
Since the 1960s, mathematicians began to use computers to help discover laws and formulate conjectures, but artificial intelligence systems have not yet been widely used in the field of theoretical mathematics.
This time, the DeepMind team and mathematicians have established a machine learning framework to assist in mathematical research.
The team also stated that their framework can encourage further cooperation between mathematics and artificial intelligence in the future.
Establish 17 test areas
Will lead and drive the innovation and development of China's artificial intelligence
On December 7, the official website of the Ministry of Science and Technology announced three letters to support the construction of the national new-generation artificial intelligence innovation development pilot zone in Harbin, Shenyang, and Zhengzhou. The three letters respectively put forward corresponding construction requirements for Harbin, Shenyang, and Zhengzhou.
The construction of the Harbin Pilot Zone should give full play to the important role of artificial intelligence in enabling the high-quality development of Harbin and the comprehensive revitalization of the old industrial base in Northeast China.
At the same time, we will give full play to Harbin's advantages of rich scientific and educational resources, distinctive industrial characteristics, and a sound foundation for international cooperation, strengthen the research and development of basic cutting-edge theories and key core technologies of artificial intelligence, and create innovative application benchmarks in fields such as smart agriculture, smart manufacturing, and cold climate scenarios.
The construction of the Shenyang Experimental Zone should give full play to the radiating and leading role of artificial intelligence in the transformation and upgrading of Shenyang's manufacturing industry and the comprehensive revitalization of the old industrial base in Northeast China, strengthen technology research and development and innovative applications, and strengthen the intelligent technology industry cluster.
Zhengzhou must give full play to the leading role of artificial intelligence in Zhengzhou's construction of a national central city, and strongly support the rise of the central region, the ecological protection and high-quality development of the Yellow River Basin.
Since Beijing became the country’s first national new-generation artificial intelligence innovation development pilot zone in 2019, Shanghai, Tianjin, Shenzhen, Hangzhou, Hefei, Deqing County, Chongqing, Chengdu, Xi’an, Jinan, Guangzhou City, Wuhan City, Suzhou City, Changsha City, Zhengzhou City, and Shenyang City have been selected successively. At present, my country has 17 national new-generation artificial intelligence innovation development pilot zones.
Our reporter Cui Shuang