Author | Victor

Edit | Qingmu

On December 9, 2021, the 6th Global Artificial Intelligence and Robotics Conference (GAIR 2021), which was jointly organized by the Guangdong -Hong Kong -Macao Greater Bay Area Artificial Intelligence and Robotics Federation and, officially opened in Shenzhen. Leaders and 30 Fellow gathered, cut in from AI technology, products, industries, humanities, organizations and other dimensions, with rational analysis and perceptual insights as the axis, and climbing the top of the wave of artificial intelligence and digitalization.

The next day of the conference, the director of the Sire Labs, the executive dean of the former Shenzhen Artificial Intelligence and Robotics Research Institute, the Academician of the International Eurasian Academy of Sciences, and IEEE FELLOW Li Shipeng gave a speech at the GAIR conference “Thinking of the Thinking of the Frontier Research on Artificial Intelligence and Robotics” Essence

Dr. Li Shipeng, IEEE FELLOW, Academician of International Eurasian Academy of Sciences. He has successively served as chief scientist and executive dean of Shenzhen Artificial Intelligence and Robotics Research Institute, vice president of Xunfei Group, and co -dean of Xunfei Research Institute, founding member and vice president of Microsoft Asia Research Institute. Academician Li is influential in multimedia, IoT and AI. He has 203 U.S. patents and publishes more than 330 papers (H index: 82). It was listed as one of the world’s top 1,000 computer scientists by Guide2research. Cultivate the winners of the four MIT TR35 Innovation Awards. It is one of the new generation of artificial intelligence industry technology innovation strategic alliance and joint secretary -general.

In the speech, Li Shipeng introduced and looked forward to the research direction of artificial intelligence and the frontier of robots. He pointed out that in the future, the data bottleneck of machine learning breakthrough deep learning may be made through the method of cognitive science. Relying on “large rules”; human -machine collaboration must also evolve into human -machine “harmony”. Only by incorporating the goals such as coupling, interaction, enhancement, and complementarity can we achieve seamless connections of human -machine.

The following is the full text of the speech.

Today’s speech topic is “Thinking of the Study on the Frontier Research of Artificial Intelligence and Robotics”, which is divided into three parts. Let’s talk about

artificial intelligence


Robot research panorama


Focus on the research direction

Including machine learning, sports intelligence, human -machine harmony, group collaboration; finally summarized.

There are three key elements of artificial intelligence related research:

Human, robot/Internet of Things and AI

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

Essence The reason why robots and the Internet of Things are classified as the interface of the two physical and virtual worlds. If there is a connection between the three elements and two elements, a new discipline will be formed, such as the combination of robots and AI will produce smart objects. Combination of AI and humans will generate human -machine coupling and enhance intelligence. Form an enhanced body. With the development of artificial intelligence and robotics, the research objects are no longer limited to single smart parties, but more and more cooperatives of multiple intelligences have been studied. For example, how can human social groups better integrate? How to design a machine group that can collaborate?

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

Overall, I think the important basic research direction is:

Machine learning, sports intelligence, human -machine harmony, group collaboration.

1 Focus on machine learning from the direction

The development of machine learning is inseparable from deep learning blessings. It has brought many research results to the industry, and empowers voice recognition, face recognition, object recognition, and autonomous driving to promote the rapid development of the artificial intelligence industry.

Although the results are quite abundant, Xiao He is also Xiao He. Deep learning depends on big data, and its bottleneck lies in big data. For example, although domestic intelligent voice technology is leading the industry, it still relies on technology accumulation and data accumulation. At present, if you want to make deep learning great power, you still need a large amount of data blessing. If you want to expand deep learning from one field to another, data support is indispensable.

How to break through? Researchers have explored multiple paths. One of the solutions is:

Expanded deep learning framework.

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

For example, optimize deep learning algorithms, knowledge map+deep learning, expert system+deep learning, etc. The other path is

Cause and effect

The goal is to use the ability of humans to fight against each other, expect to surpass the correlation between data, and then explore the causality between data, so as to obtain logical reasoning between data.

The third path is

Brain -like calculation

From the perspective of biology, explore human brain cognitive elements and mechanisms, and reproduce the human brain by simulation.

Personally, I think that cognitive science is the focus of breakthroughs in the deep learning framework. The reason is that there are two points of human cognition. We need to learn more about it: know, learn from it.

It is known that some of the cognitive ability is born, and the brain nerves of the newborn have many innate connections. The revelation it gives us is that most of the current deep learning algorithms, most of which are trained from scratch without sufficient or efficiently using prior knowledge or existing models. How to use “existing knowledge” is the next popular direction of deep learning.

It is learned that most cognitive skills are learned the day after tomorrow, especially early learning. Establish more connections by learning brain nerves. Children have many abilities, including perception, response, language, reading, writing, and understanding, and even the ideas and ability to analyze problems and solve problems. They have basically been determined when they are young; they will basically accumulate knowledge in the future. This means that when the brain nerve element is very early, it is connected to a meta -model, and the rest is just using this meta model to solve the problem of specific fields. This is amazing similar to the current large -scale pre -training model.

Another level of learning is: human learning processes rely on multi -source, multi -sensitive, multi -mode, multi -angle data, such as visual, hearing, smell, touch, and context. Most of the deep learning relies on a voice and a photo. Therefore, the input data of the AI ​​model in the future may not only be a single data, but the fusion of multiple signal sources. How to imitate the process of human learning is another revelation of cognitive science for deep learning.

Furthermore, the process of human learning is a process from the principle of sample examples to the principle, not only at the sample example level; at present deep learning is at the sample level. So, will the machine learning framework that can be constructed in the future, no matter what kind of data input, as long as the logic is connected, it will converge to the consistent model?

Breaking through the data bottleneck of deep learning, you can try to build a rules of crowdsourcing system to allow human teaching machine learning processes. The purpose is not to enter data, but to allow machine learning rules. Since we try to learn the rules from daily activities, ordinary people can mark the teaching of such rules, which breaks the limitations of “experts” in the past. This method of transition from “big data” to the “large -scale” model is obviously more in line with human cognition.

2 Focus on the movement of the direction of the direction

As we all know, in the field of robotics, Boston Power’s products are the most “like people”. As in the upper map, robots dancing can not see a hard feeling. But due to the restrictions on computing resources, energy, and motion control, it can only run for dozens of minutes. In fact, the operation of Boston power robots is based on motor -based drivers. There are many disadvantages, such as the contradictions of rigid motion, large self -weight, response speed and flexibility, and large energy consumption.

Compared with the operation of humans and other animals, the combination of muscles, bones, sensing and nerves can achieve flexible operation in low energy consumption. This enlightenment to researchers is that the operating system of the robot should be as satisfying as people: efficient, flexible, accurate, robust, rigidity, softness, lightweight, adaptive and other indicators. The current exercise intelligence may perform well in a certain dimension, but there are still many shortcomings in comprehensive considerations.

Therefore, an important research direction of sports intelligence is: bionic. The motion intelligence of animals, such as exercise control adopts approximation of feedback, and flexibly adjust the visual changes in the movement process at any time.

If the robot is driven by “internal force”, medical micro -nano robots are the representative of the “external force” research direction. For example, relying on magnetic force, small robots accurately transport the drug from one pipe to another.

3 Focus on the human -machine harmony

At the level of human -machine harmony, it is different from collaboration. “Harmony” represents the meaning of coupling, interaction, enhancement, complementarity, collaboration, and harmony in the collaboration of human -machine collaboration. The goal of human -machine harmony is: no need to tell the human intention of human beings, the machine can understand, so as to achieve seamless connection of human -machine.

In the process of reaching the harmony of human machines, focusing on the natural interaction, perception and enhancement of human -computer. Specific may include: biological characteristics detection and recognition, human -machine interface, brain interface, voice recognition, action recognition, expression recognition, language understanding, intention understanding, body perception, gap enhancement, and extension of extension and remote reality, etc. Wait.

In terms of human -machine enhancement intelligence, most of today’s machine learning frameworks are based on big data deep learning frameworks. They will definitely encounter a scene that machine intelligence cannot handle. This is fatal for certain high -risk areas, such as autonomous driving and finance.

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

In response to this problem, the current solution is “human takeover”. This will involve three core issues:

Core question 1: How can machine intelligence feel that they can’t handle some situations, and take the initiative to ask people to take over?

Core question 2: When can humans completely let go of the machine to complete the task independently?

Core question 3: What kind of human -machine interaction design can give full play to the strengths of people and machines, while at the same time, do you need to trouble the other party?

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

If the three core issues cannot be resolved, it will lead to some difficulties. For example, taking autonomous driving as an example, the current safety officer does not open the “automatic” function once and for all. It still needs to monitor road conditions and routes from time to time. This actually increases the burden on the safety officer, because when there is no autonomous driving, humans will have certain predictions on their own driving environment, and human driving is unpredictable.

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

The human -machine enhancement body is also a field of human -machine harmony, which can help humans enhance physical body capabilities and complete some things that humans cannot finish their physical strength. However, the machine may be too complicated and requires human training to operate. The future goal of human -machine enhancement body is to achieve harmonious coexistence of man and machine, and to control it as natural as human organs. Among them, the core research topics involved include: the intention of the machine, the attitude of the person, the understanding of the natural language command of the person, the body language, etc., so that the machine can help people solve the problem in a smooth way to accept the smooth and just smooth way.

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

4 Focus on group collaboration

At present, single intelligence can already complete many tasks, but how to use the power of each smart body? This involves the research direction of group collaboration. In the warehousing scenario, there are many robots that capture classifications. If it can be effectively scheduled, it will greatly improve work efficiency.

The current mainstream scheduling method is a centralized control method, but in the face of thousands of intelligence of intelligence, non -center control is required, allowing autonomous behaviors to allow smart parties. “Do your own thing.” That is, smart and independent operations with intelligence, more efficient group/system intelligence and behavior through collaboration.

The current rules involved in smart group collaboration include group behavior models and incentive mechanisms, and group intelligent collaborative decisions. In this regard, ants are our learning objects. In addition, in terms of autonomous driving, more and more independent driving robots appear. How to achieve collaborative perception and collaborative control is also today.

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

The above four aspects are basic research, and a breakthrough in any field will be a revolutionary breakthrough for its fields and downstream applications. Occupy the advantage in competition!

The first day of GAIR 2021: The 40 -year AI years of 18 -bit Fellow, a technical frontier inheritance and excitement

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考


Tourism legend: China has handled it for forty years in parallel, and they explored the Golden Age of Calculating from the Unmanned Area | GAIR 2021


IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考

Power of Time -1991 Artificial Intelligence Debate 30th Anniversary: ​​Doctrine is no longer.


On Sanyi, connecting communication in series, penetrating the edge, evolving cognition, and remitting the machine: Listen to five IEEE FELLOW talk about the future of AI | GAIR 2021

IEEE Fellow李世鹏:人工智能与机器人前沿研究之思考


Where does the new generation of AI talents come from and where should I go? | GAIR 2021 Dean Forum