IEEEFellow Li Shipeng: Thoughts on the Frontier Research of Artificial Intelligence and Robots

Edit | Mu Qing

On February 9th, 202 1, 1, the 6th Global Congress of Artificial Intelligence and Robotics was jointly organized by Guangdong-Hong Kong-Macao Greater Bay Area Federation of Artificial Intelligence and Robotics and Lei Feng. Com, officially kicked off in Shenzhen. 140 More than 30 industry and academic leaders and Fellow gathered together, starting from the dimensions of AI technology, products, industry, humanities and organization, with rational analysis and perceptual insight as the axis.

On the second day of the conference, Li Shipeng, director of Sill Laboratory, former executive dean of Shenzhen Institute of Artificial Intelligence and Robotics, member of the International Eurasian Academy of Sciences and IEEEFellow, gave a speech on "Thinking about the Frontier Research of Artificial Intelligence and Robotics" at the GAIR conference.

Dr. Li Shipeng, IEEEFellow, Member of the International Eurasian Academy of Sciences. He has served as chief scientist and executive dean of Shenzhen Institute of Artificial Intelligence and Robotics, vice president of Iflytek Group, co-president of Iflytek Research Institute, founding member and vice president of Microsoft Research Asia. Academician Li is influential in the fields of multimedia, Internet of Things and artificial intelligence. He holds 203 American patents and has published more than 330 cited papers. It is listed as one of the top 1000 computer scientists in the world by Guide2Research. Four MITTR35 Innovation Award winners received training. He is one of the founders and co-secretary-general of the strategic alliance for technological innovation in the new generation of artificial intelligence industry.

In his speech, Li Shipeng introduced and prospected the frontier research direction of artificial intelligence and robotics. He pointed out that in the future, machine learning can break through the data bottleneck of deep learning, and the learning paradigm can be changed from relying on big data to relying on big rules; Man-machine cooperation should also evolve into man-machine harmony. Only by bringing the goals of coupling, interaction, enhancement and complementarity into the research direction can the seamless connection between man and machine be realized.

The following is the full text of the speech, and the AI ? ? technical review has been sorted out without changing its original intention:

The topic of today's speech is "Thinking about the Frontier Research of Artificial Intelligence and Robots", which is divided into three parts. Let's talk about it first

artificial intelligence

and

Panorama of robot research

, and then

Focus on the research direction

, including machine learning, sports intelligence, man-machine harmony and group cooperation; Finally, make a summary.

There are three key elements in the research of artificial intelligence:

People, Robots/Internet of Things and Artificial Intelligence

. Robots and the Internet of Things are grouped together because they are the interface between the physical world and the virtual world. If these three elements are interrelated, a new discipline will be formed. For example, the combination of robot and AI will produce agent, the combination of AI and human will produce man-machine coupling and enhance intelligence, while the integration of robot and human will form reinforcement. With the development of artificial intelligence and robotics, the research object is no longer confined to a single agent, but more and more research is focused on the cooperation of multi-agents, such as how to better integrate human social groups. How to design a set of machines that can cooperate skillfully?

Generally speaking, I think the important basic research direction is:

Machine learning, sports intelligence, man-machine harmony, group cooperation.

1

Machine learning of focusing direction

The development of machine learning is inseparable from the blessing of deep learning, which brings many research results to the industry, empowers speech recognition, face recognition, object recognition, autonomous driving and other aspects, and promotes the rapid development of artificial intelligence industry.

Despite the fruitful results, Xiao He failed to succeed. Deep learning depends on big data, and its bottleneck also lies in big data. For example, although domestic intelligent voice technology is in the leading position in the industry, it still relies on technology accumulation and data accumulation. Now, if you want deep learning to exert great power, you need a lot of data blessing. If you want to expand deep learning from one field to another, you need data support.

How to break through? Researchers have explored many ways, and one of the solutions is:

Extended deep learning framework.

For example, optimizing deep learning algorithm, knowledge map+deep learning, expert system+deep learning and so on. The other way is

Causal reasoning

With the help of human's ability to extrapolate, it is expected to transcend the correlation between data, and then explore the causal relationship between data, so as to get logical reasoning between data.

The third way is

Brain like computing

From the perspective of biology, this paper explores the cognitive elements and mechanisms of the human brain, and reproduces the human brain through simulation.

Personally, cognitive science is the key to break through the deep learning framework. The reason is that there are two points in the process of human cognition that we need to learn from: being born with knowledge and learning with knowledge.

Being born knowing means that some cognitive abilities are innate, and there are many innate connections in the newborn's brain. What it gives us is that most of the current deep learning algorithms are training from scratch, without fully or efficiently using prior knowledge or existing models. How to use the existing knowledge is the next hot direction of deep learning.

Learning to know means that most cognitive abilities are acquired, especially early learning. By studying the cranial nerves, more connections have been established. Many children's abilities, including perception, coping, language, reading, writing, understanding, and even the thinking and ability to analyze and solve problems, have been basically stereotyped at an early age; It is basically the accumulation of knowledge in the future. This means that brain neurons have been connected and shaped into a metamodel for a long time, and the rest is just to use this metamodel to solve problems in specific fields. This is strikingly similar to the current large-scale pre-training mode.

Another level of learning is that human learning process depends on multi-source, multi-sensor, multi-modal and multi-angle data, such as joint information of vision, hearing, smell, touch and context. Today's deep learning mostly depends on a voice and a photo. Therefore, the input data of AI model in the future may not only be a single data, but a fusion of multiple signal sources. How to imitate the process of human learning is another enlightenment of cognitive science to deep learning.

Furthermore, human learning process is a process from sample examples to principle induction, not just at the level of sample examples; At present, deep learning is only at the sample level. Then, can we construct a humanoid machine learning framework in the future? No matter what kind of data is input, as long as the logic is consistent, it will converge to a consistent model?

To break through the data bottleneck of deep learning, we can try to build a regular crowdsourcing system, so that human beings can teach the process of machine learning, and its purpose is not to input data, but to let machines learn rules. Because we try to learn rules from daily activities, such rules can be marked and taught by ordinary people, breaking the limitation that experts need expert systematization in the past. This transition from big data to big rule model is obviously more in line with human cognition.

2

Motion intelligence in focusing direction

As we all know, in the field of robotics, the products of Boston Dynamics are the most like people. As shown in the above picture, robots can't see the feeling of being born hard when dancing. But limited by computing resources, energy and motion control, it can only run for tens of minutes. The operation mode of Boston power robot is actually based on motor drive, which has some shortcomings, such as high rigidity, heavy weight, contradiction between reaction speed and flexibility, and high energy consumption.

Compared with the running modes of humans and other animals, the combination of muscles, bones, sensors and nerves can realize flexible operation with low energy consumption. The enlightenment to researchers is that the operating system of robots should meet human requirements: efficient, flexible, accurate, robust, flexible, portable and adaptable. Nowadays, sports intelligence may perform well in a certain dimension, but there are still many shortcomings in comprehensive consideration.

Therefore, bionic is an important research direction of sports intelligence. Imitating animal's sports intelligence, such as adopting approximate feedback in motion control, can flexibly adjust the motion process at any time according to changes.

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

three

Man-machine harmony in focusing direction

On the level of man-machine harmony, it is different from cooperation. Harmony represents coupling, interaction, enhancement, complementarity, cooperation and harmony in man-machine cooperation. The goal of man-machine harmony is that the machine can understand without telling the human intention, so as to realize the seamless connection between man and machine.

In the process of realizing man-machine harmony, we pay attention to the natural interaction, perception and enhancement between man and machine. Specifically, it may include: biometric identification, human-computer interface, brain-computer interface, speech recognition, action recognition, expression recognition, language understanding, intention understanding, posture perception, seamless enhancement, and the extension of augmented reality and remote reality.

In terms of man-machine augmented intelligence, most of the current machine learning frameworks are deep learning frameworks based on big data, and there will definitely be situations that machine intelligence cannot handle. This is fatal for some high-risk areas, such as autonomous driving and finance.

To solve this problem, the current solution is manual takeover. This will involve three core issues:

Core question 1: How does machine intelligence perceive that it can't handle some situations and ask someone to take over?

Core question 2: When can humans completely let machines complete tasks autonomously?

Core question 3: What kind of human-computer interaction design can give full play to the respective strengths of man and machine without unnecessarily disturbing each other?

If the three core problems cannot be solved, it will lead to some difficulties. For example, taking automatic driving as an example, at present, the safety officer does not turn on the automatic function once and for all, and still needs to monitor the road conditions and routes from time to time without being distracted for a moment. This actually increases the burden on security personnel, because when there is no automatic driving, human beings will have certain predictions about their driving environment, while human beings can't predict the situation of machine driving.

Man-machine enhancement of the body also belongs to the field of man-machine harmony, which can help human beings to enhance their physical ability and accomplish something that human beings cannot accomplish by their own physical strength. But the machine may be too complicated and need manual training to operate. The future goal of man-machine augmented organisms is to realize the harmony between man and machine, just as natural as human organs. Among them, the core research topics involved include: machine perception of human intentions, human gestures, understanding human natural language commands, body language and so on. So that machines can help people solve problems in a smooth way that is most suitable for human acceptance.

four

Focus on the direction of group cooperation

At present, a single agent can accomplish many tasks, but how to exert the power of each agent? This involves the research direction of group cooperation. In the warehouse scene, there are many robots that grab and sort. If they can be effectively scheduled, the work efficiency will be greatly improved.

At present, the mainstream scheduling mode is centralized control mode, but in the face of thousands of large-scale agents, it is necessary to decentralize control, so that agents can cooperate with each other and have independent behavior and do their own things well. In other words, a single intelligent and independent agent can achieve more efficient group/system intelligence and behavior through cooperation.

At present, the rules involved in agent group cooperation include group behavior model and incentive mechanism, and ants are our learning objects in group intelligent collaborative decision-making. In addition, in the aspect of autonomous driving, there are more and more autonomous robots, and how to realize cooperative sensing and cooperative control between them is also a hot topic today.

The above four aspects belong to basic research. Any breakthrough in any field will be a revolutionary breakthrough in its field and downstream application, and it will also bring innovation in the original technology of industrial digital intelligence, which will give us an advantage in the competition!

Leifeng. com