In 2022, artificial intelligence will bring more surprises to human beings
◎Intern reporter Du Peng
2022, which is about to pass, is a year worth remembering for artificial intelligence.
A large number of artificial intelligence-related applications have stepped out of the laboratory and are constantly moving towards large-scale implementation.
The Beijing Winter Olympics under the blessing of AI "black technology" is brilliant; unmanned driving has opened multi-city pilots, and the future transportation will go further; AI paintings are fascinating, and artistic creation may no longer be exclusive to humans...
Whether it is the continuous breakthrough of the underlying technology or the flourishing of various applications, in the past year, artificial intelligence has shown us its infinite possibilities.
We believe that this is just the tip of the iceberg of artificial intelligence, and it has more potential waiting for us to tap in the future.
With the continuous maturity of technology and continuous innovation of landing applications, artificial intelligence may truly change your life and mine.
AI "black technology" illuminates the Beijing Winter Olympics
Helps weather forecast, game broadcast and sign language broadcast, etc.
On February 4, the 2022 Beijing Winter Olympics, which attracted worldwide attention, officially kicked off.
The application of artificial intelligence and other technologies has added a different kind of "beauty of science and technology" to this Winter Olympics.
At the Winter Olympics, the artificial intelligence MOML algorithm developed by Zhang Pingwen, academician of the Chinese Academy of Sciences, vice president of Peking University, and chief scientist of Chongqing Big Data Research Institute of Peking University, empowered the weather forecast model, making the weather forecast for the Winter Olympics more accurate.
The inherent advantages of artificial intelligence algorithms in the fusion and processing of information make them able to replace forecasters to a certain extent in information integration and analysis in consultations, and through data mining and learning, the experience of forecasters can be internalized in the algorithm. While improving the efficiency of weather forecasting, it also further improves the accuracy of forecasting.
In the women's freestyle skiing final of this Winter Olympics, Chinese player Gu Ailing won her first personal gold medal with her wonderful performance against the sky.
During the game broadcasting process, Baidu Smart Cloud uses the "3D+AI" technology to create the "same field competition" system, which turns a single-player event into a "multiplayer game", realizing the three-dimensional restoration and virtualization of the championship and runner-up games. Superimposition is convenient for the audience to see the real-time movements of different players; at the same time, quantitative analysis of the athletes' movements is carried out through technical means, and a series of motion data such as skating speed, flying height, landing distance, and rotation angle are superimposed on the original picture, so that the audience You can more intuitively understand the differences in technical movements between players from the perspectives of fluency, completion, difficulty, diversity and aesthetics.
On the same day as the opening of the Beijing Winter Olympics, the AI sign language anchor of CCTV News also officially took up her post. She provided real-time sign language interpretation services for the hearing-impaired during the Winter Olympics news broadcast, live broadcast of events and live interviews.
With an accurate sign language translation engine, the AI sign language anchor has an intelligibility of more than 85%, and can quickly and accurately convert text, audio and video content of ice and snow events into sign language.
Tencent's "Hunyuan" AI model topped the VCR list
Demonstrated its strong strength in the field of multimodal understanding
On May 31, Tencent's "Hunyuan" AI large model reached the top of the VCR (Visual Commonsense Reasoning, Visual Commonsense Reasoning), an international authoritative list in the field of multimodal understanding, ranking first in both individual scores and total scores.
This is another major breakthrough of the "Hunyuan" AI large model after the Grand Slam in the field of cross-modal retrieval, the CLUE natural language understanding classification list and the CLUE general list, showing its strength in the field of multi-modal understanding. strong strength.
Different from the cross-modal understanding task, the multi-modal understanding task requires the computer not only to be able to achieve recognition-level perception (such as classification detection, etc.), but also to achieve cognitive-level perception (such as judgment intention, logical reasoning, etc.).
The "Hunyuan" AI large model that topped the VCR list this time was independently developed by the Tencent Advertising Multimedia AI team. At the same time, it carried out pre-training tasks and training methods with the help of the graphics processor computing power and training acceleration framework of the Tencent Taiji machine learning platform. Many innovative improvements and designs have been made to effectively improve the performance of the model.
Up to now, the "Hunyuan" AI large model has won the first place in the authoritative list of AI in many fields such as MSR-VTT, MSVD, CLUE, and VCR, and has set a number of historical records in the industry.
This means that the technical strength of "Hunyuan" in the fields of natural language understanding, multi-modal understanding, and cross-modal understanding has been verified.
Google engineer makes an oolong claim that AI has consciousness
The so-called "personality" of artificial intelligence is more like imitating human beings
Google's AI engineer made a fuss, claiming that the LaMDA language model is conscious, triggering a discussion in the industry on "whether AI has autonomous consciousness".
In June of this year, Lemoine, an AI engineer at Google, believed that the dialogue application language model LaMDA has "self-awareness", and issued 21 pages of evidence for this.
Lemoine believes that there are three reasons why LaMDA is conscious: one is that LaMDA uses language efficiently and creatively than ever before; two is that it shares sensations in a similar way to humans; Worried about the future, but also reminiscing about the past.
LaMDA is a large-scale natural language dialogue model announced by Google at the 2021 Developer Conference. It can simulate any entity with knowledge attributes, and answer questions for users in a friendly and natural dialogue with humans through an "anthropomorphic" approach. Pass on more knowledge.
Lemoine's arguments and evidence have drawn widespread attention in the industry.
Shortly after, Google issued a statement saying Lemo was fired for violating its "Employment and Data Security Policy."
Google said that after an extensive review, they found Lemoine's claims that LaMDA was alive to be completely unfounded.
Experts generally believe that the so-called "personality" of current artificial intelligence is more of an imitation of human language style. AI with self-awareness and perception should be active and have a unique perspective on people and things. It's just a computer system designed by people as a tool to do certain things.
The world's first three-modal large-scale model of graphics, text and audio was born
"Zidong Taichu" realizes "sounds from pictures" and "pictures from sounds"
On September 1, at the 2022 World Artificial Intelligence Conference held in Shanghai, the "Zidong Taichu" multi-modal large-scale model project jointly developed by Wuhan Institute of Artificial Intelligence, Institute of Automation, Chinese Academy of Sciences, and Huawei Technologies Co., Ltd. won the award. The highest award of the conference.
"Zidong Taichu" is the world's first three-modal large-scale model of graphics, text, and audio. It pioneered the "unified representation" and "mutual generation" of image, text, and voice three-modal data. The ability to understand and generate images is closer to that of humans, providing an innovative foundation for creating multi-modal artificial intelligence industry applications, and taking an important step towards general artificial intelligence.
The core principle of the mutual transformation and generation among the three modalities of "Zidong Taichu" is that the different modalities of vision, text, and voice are mapped to a unified semantic space through their respective encoders, and then learn the relationship between the modalities through a multi-head self-attention mechanism. Semantic association and feature alignment form a multi-modal unified knowledge representation; after that, the encoded multi-modal features are used to generate text, images and voices respectively through the decoder.
With four major breakthroughs, "Zidong Taichu" effectively helps the development of general artificial intelligence with multi-modal cognition as the core.
First, it proposes a multi-level, multi-task cross-modal self-supervised learning framework for the first time, supporting a three-level pre-training self-supervised learning method from the entry level to the modality level and sample level; second, it completes the semantics of weakly correlated multi-modal data for the first time Unified representation, reducing the cost of data collection and cleaning; the third is the first unified modeling of multi-modal understanding and generation tasks, supporting cross-modal retrieval, multi-modal classification, speech recognition, image generation and other understanding and generation tasks; the fourth is It is the first time to achieve unsupervised and surpass supervised methods, based on 5%-10% data annotation, to achieve 100% supervised learning effect.
AI breaks matrix multiplication calculation speed record
Solved a 50-year-old problem in mathematics
In October, the cover of the British "Nature" magazine was titled "Matrix Game", and published the latest discovery of the artificial intelligence company "Deep Thinking" team: AI can solve the matrix multiplication problem.
The AI system, dubbed "AlphaTensor," discovers new algorithms on its own, solving a 50-year-old unsolved problem in mathematics -- finding the fastest way to multiply two matrices together.
This is the first AI system to discover novel, efficient and correct algorithms for basic tasks such as matrix multiplication.
Mathematics appears frequently in computer programming, often as a means of describing and manipulating representations of real-world phenomena.
For example, it can be used to represent pixels on a computer screen, weather conditions, or nodes in an artificial network.
One of the main ways to use mathematics in this context is to perform calculations on matrices.
The larger the matrix, the greater the workload, and computer scientists began to spend a lot of time and energy developing more efficient algorithms to complete the related work.
In this latest effort, DeepMind researchers explored whether it is possible to use reinforcement learning-based AI systems to create new algorithms that require fewer computational steps than existing algorithms.
To find out, they looked to game systems for inspiration.
After building some preliminary systems, the research team turned its focus to tree search, a way for the system to look at various scenarios in a given situation.
Next, the researchers will allow the system to create its own algorithms, further improving efficiency.
They found that, in many cases, the algorithms chosen by the system were better than those created by humans.
The "Deep Mind" team hopes that in the future, AI will be used more to help overcome some important problems in the fields of mathematics and science.
2022 China Artificial Intelligence Innovation and Development Index Announced
Fully reflect the development trend of artificial intelligence in my country
On November 18, at the opening ceremony of the 5th World Sound Expo and the 2022 HKUST Xunfei Global 1024 Developers Festival, China Electronics Information Industry Development Research Institute (also known as CCID Research Institute) released the 2022 China Artificial Intelligence Innovation and Development Index (Hefei Index).
This is the first national artificial intelligence special research achievement named after a region in my country, aiming to comprehensively and systematically reflect the development trend of artificial intelligence in China.
The China Electronics Information Industry Development Research Institute has constructed the China Artificial Intelligence Innovation and Development Index, which is the evaluation system of the "Hefei Index", from five dimensions: development environment, innovation capability, infrastructure support, capital investment, and industrial strength.
In recent years, my country's artificial intelligence has entered a new stage of deep integration and application with the economy, intelligent transformation has been comprehensively promoted, and the global influence of the artificial intelligence industry has continued to increase.
In 2021, the research and development intensity of artificial intelligence in my country will be 19.4%, and the number of employees will increase to 310,000, accounting for 5.3% of the global proportion.
From 2017 to 2021, the scale of my country's artificial intelligence industry has increased by 2.6 times, accounting for 16.8% of the world's total.
The proportion of patent applications in the world continues to expand, from 13% in 2012 to 70.9% in 2021.
In terms of innovation capabilities, my country's artificial intelligence research and development investment has continued to increase, and the number of employees has continued to increase.
From the perspective of the overall index, Beijing, Guangdong, and Shanghai are in the leading positions in the field of artificial intelligence, and Anhui is closely behind, ranking sixth in the country.
Hefei has become one of the most active cities in the field of artificial intelligence, technological innovation and industrial development.
ESMFold predicts more than 600 million protein structures
Prediction speed is 60 times faster than "alpha folding"
The British "Deep Thinking" company announced in August that its artificial intelligence program "Alpha Fold" has predicted more than 200 million protein structures from about 1 million species, covering almost every protein structure that has been cataloged in the scientific community .
But just this November, researchers at Metaverse Platforms (Meta) used the artificial intelligence model ESMFold to predict the structures of more than 600 million proteins from bacteria, viruses, and other as yet uncharacterized microbes.
In this latest study, the research team used large-scale language models to predict these protein structures.
It is reported that language models usually require a large amount of text for training. In order to apply this model to protein structure prediction, the research team used known protein sequences to train it. These known proteins can be expressed by chains composed of 20 different amino acids. Each amino acid is represented by a letter.
ESMFold then learned to "autocomplete" protein structure predictions with ambiguous amino acid ratios.
These trainings gave ESMFold an intuition about protein sequences that contain information about their shape, says team leader Alexander Reeves.
And, like Alpha Folding, the model combines this learned information with information about relationships between known protein structures and sequences to generate predicted structures.
The team pointed out that although ESMFold's predictions are not as accurate as "alpha folding", they are 60 times faster in prediction speed, which means that it can expand the structure prediction database to a larger scale.
The first AI modeling method for protein dynamic structure
It is of great significance for understanding life processes and developing new drugs
On December 8, Westlake University announced the AI model that can describe protein conformational changes and affinity predictions——ProtMD, which was first developed by the team of Li Ziqing, chair professor of artificial intelligence at the school, together with Xiamen University and Hangzhou Derui Zhiyao Technology Co., Ltd.
This is the first artificial intelligence model that attempts to analyze the dynamic conformation of proteins, which can assist medicinal chemists to more accurately screen out highly active small molecules, thereby accelerating preclinical drug development.
Previously, the "Alpha Fold 2" developed by Google's company can use artificial intelligence to accurately predict the three-dimensional structure of proteins, which has had a huge impact on structural biology, drug design and even the entire scientific community.
However, "Alpha Fold 2" can only predict the static structure of a protein at an instant, and has not been able to solve the prediction of dynamic changes in protein structure.
The AI model developed by Li Ziqing's team this time can predict the change process of the protein structure after the drug molecule binds to the target protein in the body (flexible docking), and infers the drug and target protein given the drug molecule and the target protein. The stability of the combination can predict the drug function, thereby improving the accuracy and efficiency of AI drug design.
Li Ziqing said that predicting the dynamic changes of protein structures is of great significance for understanding life processes and developing new drugs.
Especially in AI drug design, it is an important idea to improve the accuracy and efficiency of AI drug screening by predicting the dynamic structural changes after the drug molecule binds to the target protein, and evaluating the drug-target binding affinity and drug effect.
Multiple cities promote the development of the autonomous driving industry
my country's autonomous driving industry is officially moving towards L3
2022 is a milestone year for the autonomous driving industry. Relevant policies have been introduced intensively, and related applications have moved from R&D testing to large-scale commercial pilots.
At present, nearly 30 cities across the country have issued more than 1,000 road test licenses to more than 80 companies, allowing high-level intelligent networked vehicles to carry out large-scale test demonstrations of carrying people and objects in specific scenarios and special areas.
More and more cities are advancing the commercialization of higher levels of autonomous driving.
On August 1 this year, the "Regulations on the Administration of Intelligent Connected Vehicles in Shenzhen Special Economic Zone" came into effect. The regulations propose that L3-level autonomous driving will be open to road testing and demonstration applications in the entire administrative region, and explore the pilot commercial operation, marking my country's autonomous driving industry. Officially moving towards the L3 level.
Since then, government departments in Chongqing, Wuhan and other places have successively issued pilot policies for fully unmanned commercialization of autonomous driving, and issued Baidu the first batch of unmanned demonstration operation qualifications in the country, allowing autonomous vehicles without safety personnel in the car to operate on social roads. to develop commercial services.
In addition, in order to promote the healthy and orderly development of the intelligent networked vehicle industry, the Ministry of Industry and Information Technology and the Ministry of Public Security also organized the drafting of the "Notice on Carrying out the Pilot Work of Intelligent Networked Vehicles Access and Road Passage (Draft for Comment)". Select qualified road motor vehicle manufacturers and intelligent networked vehicle products equipped with automatic driving functions that are qualified for mass production, and carry out access pilots; for intelligent networked vehicle products that have passed the access pilot program, limited public roads in pilot cities Carry out on-road traffic pilots in the region.
AI painting is on fire, and the first year of AIGC starts
It is expected to generate trillions of economic value in the future
In August of this year, in the Emerging Digital Artist Competition held in Colorado, USA, the AIGC painting submitted by contestant Jason Allen - "Space Opera House" won the category of "Digital Art/Digital Retouched Photo" of the competition first prize.
Jason Allen, who has no drawing foundation, borrowed an AI drawing tool called Midjourney. Through a process similar to a "word game", after entering keywords related to the picture effect, such as subject matter, light, scene, angle, atmosphere, etc., I got the initial works, and finally completed this group of "Space Opera House" digital art works after repeated adjustments and modifications.
This year, AI drawing applets, websites, etc. began to grow rapidly, and software such as Meitu Xiuxiu and Douyin also added AI drawing functions.
According to data from the Douyin platform, as of December 6, more than 24.284 million people have used the special effect, and it quickly soared to the top of the special effect trend list.
The Baidu index of AI painting has also risen from an average of 2,000 to 3,000 per day to 30,000 per day, which shows how popular it is.
The popularity of AI painting has also gradually brought the concept of AIGC into the public eye.
The so-called AIGC (AI Generated Content) is a new production paradigm based on artificial intelligence technology to automatically generate content.
Its technology mainly involves two aspects: natural language processing (NLP) and AIGC generation algorithm.
Among them, natural language processing is a means to realize the interaction between humans and computers through natural language.
Initially, the content forms that AIGC can generate are mainly text. After exponential development in 2022, the content forms that can be generated by AIGC technology have been expanded to include text, images, videos, voices, codes, robot actions, etc. Therefore, 2022 is also called "the first year of AIGC".
Generative AI allows machines to start doing knowledge and creative work at scale, and is expected to generate trillions of dollars in economic value in the future.
(Science and Technology Daily)