CHEATC2023|Chen Xiaoping, University of Science and Technology of China: From large models to soft robots

Since the beginning of this year, with the continued popularity of ChatGPT, large models have also entered a period of rapid development. Many well-known domestic and foreign technology companies have successively launched independently developed large model products. So what is the technical principle of large models?

On May 18, Professor Chen Xiaoping, director of the Robotics Laboratory of the University of Science and Technology of China, who was invited to participate in the 2023 China Home Appliances Technology Conference (CHEATC2023), shared his research and views. He is also the director of the Artificial Intelligence Ethics of the China Artificial Intelligence Society. Together with the Chairman of the Governance Committee, Professor Chen Xiaoping delivered a keynote speech on "New Developments in Artificial Intelligence: From Large Models to Soft Robots" at this conference, introducing the technical principles of large artificial intelligence models and new developments in the application of artificial intelligence. Technology trends.

Professor Chen Xiaoping of University of Science and Technology of China

"The fundamental principle of big models is to make predictions," Chen Xiaoping said. The development of artificial intelligence has now begun the process of the fourth wave, and data models have also shifted from big data-driven to big training-driven. Different from the previous three waves, the new stage of artificial intelligence has new requirements for the quality, quantity and acquisition methods of training data, and finally forms an example model that can be applied to large-scale real scenes. He emphasized that a large model is an intelligent system integrated by multiple technologies, rather than a simple combination of a single or a few technologies. ”

The rise of large models comes from generative artificial intelligence. Currently, generative artificial intelligence is not just a simple generation of content such as language and images, but also completes intelligence based on the precise processing of human natural language. Human-computer interaction. Chen Xiaoping said: "At this stage, our expectation for machine language processing is that it can speak human language, understand human language, and answer questions, even if the answer may not be correct. Among them, the basic requirement is that the speech must conform to human language habits. "Since there are no scientific standards for human language habits but there are empirical standards, how can machines master and utilize human language habits? Chen Xiaoping said: "The basic research idea and secret of success of large models is: extract language from large-scale human corpora. Traces and used in human-computer natural language interaction.”

The large model extracts semantic elements including characters, words, punctuation marks, etc. from the original human corpus, and then performs semantic review based on the correlation between the preceding and following semantic elements, and finally achieves behavioral prediction. In principle, the greater the number of semantic elements that are looked back, the higher the accuracy of prediction. At least 4,000 tokens can be reviewed by large models, and some models can review up to 100,000 tokens. "Chen Xiaoping said. The large model technology system is based on the pre-trained model, and then uses a specially trained special model to cooperate with the user guidance model to accurately understand and answer the user's questions. The three major models cooperate with each other, and the quality of the artificial intelligence answer can be Achieve substantial improvement.

Although the emergence of large-scale models has brought new innovative directions to artificial intelligence, it is not applicable to all aspects of real-world scenarios. According to Chen Xiaoping, the three major areas of artificial intelligence that China currently needs to conquer are intelligent manufacturing, intelligent agriculture and inclusive elderly care. "Overcoming these three major battles will completely change our global landscape." On the other hand, large models bring huge changes but also bring new challenges. When large models are based on imitations of human functions, they are likely to be thought of as having emotions and consciousness. This is because people habitually apply their understanding of a concept to the overall structure involving that concept, thinking that the information expressed by the structure also has the same meaning, but in fact this is not the case. "Chen Xiaoping said that the application of large models may also have public safety, employment and long-term impacts.

In addition to large-scale models, Professor Chen Xiaoping also achieved new scientific research results on "artificial intelligence in the physical world". At present, the physical form of artificial intelligence we put into application is mainly rigid robots. This kind of robot has high repeatability, but low dexterity and safety. It is suitable for structured environments, but needs to be carried out in unstructured environments. Accurate measurement, modeling and calculation require high technical requirements and are currently not suitable for most industries. In response to these shortcomings of rigid robots, Chen Xiaoping proposed the principle of fusion, under the three basic assumptions that accurate measurement of the operating objects of intelligent robots is not feasible, accurate modeling of the working environment and operating objects is not feasible, and precise decision-making is not feasible. , developed a pneumatic honeycomb network software arm. This kind of arm has good performance in terms of flexibility and load capacity, and can achieve precise control when there is external interference and irregular movement of objects. It is expected that this technology will have broad application prospects in fields such as home services, emotional interaction, and autonomous driving. On the other hand, Chen Xiaoping's team also combined a flexible arm with a rigid machine, resulting in the experimental results of "rigid and soft claws in one", which enabled the robot to achieve control without changing the program and hardware parameters or using force feedback sensors. Precise gripping of multi-shaped objects.

The above is the detailed content of CHEATC2023|Chen Xiaoping, University of Science and Technology of China: From large models to soft robots. For more information, please follow other related articles on the PHP Chinese website!