Song Liang, Academician Of The Canadian Academy Of Engineering: Future Artificial Intelligence Systems Will Exist in The Form Of Networks

C114 News on December 5th. Approved by the State Council and co-sponsored by the Ministry of Science and Technology and the Henan Provincial Government, the 2023 World 5G Conference with the theme of "5G Transformation Shapes the Future" will be held in Zhengzhou City, Henan Province from December 6th to 8th. Held at Zhengzhou International Convention and Exhibition Center.

Focusing on "strengthening the foundation, toughening the chain and leading the way" and "enabling high-quality development of the industry", the 2023 World 5G Conference will set up 12 parallel forums to help the world's top 5G industry cooperation and resource integration. On the eve of the official opening of the World 5G Conference, the "Tech Talk 2023 Innovation Technology Forum" with the theme of "Integrated Innovation to Enhance Value Space" kicked off today.

During the forum, Song Liang, academician of the Canadian Academy of Engineering, professor of Fudan University, and chairman of the International Society for Intelligent Network Systems, was invited to give a keynote speech titled "From Large Models to Online Evolutionary Learning."

Song Liang said that AIGC represented by ChatGPT has ignited the industry's enthusiasm for artificial intelligence AGI, but GPT has not yet reached general artificial intelligence, and there are still two basic problems that have not been solved in large models. The first is how to evolve online? A characteristic of natural language processing is that it uses the first few words to predict what the next word will be. Humans have done a good job of organizing and annotating the data for the machine. The machine only needs to learn; but the current machine learning paradigm cannot learn. Self-taught knowledge. Secondly, the current large model data volume is too large. For general artificial intelligence, how to enable artificial intelligence to independently mine data that humans have not labeled so that it can generate new knowledge is a problem that needs to be solved.

With the development of 5G and computing networks, the physical environment will generate a lot of new data in real time. These data themselves are not labeled. How can we use distributed massive computing methods to unlabel the data generated in real time in the network and the physical world? The industry is paying great attention to the processing of online evolution. Song Liang pointed out that by building an intelligent environment for multi-agent collaboration, the physical environment, information environment and human society are coupled into an intelligent environment, so that it can supervise the independent training and evolution of each agent, thereby realizing artificial intelligence Without human intervention, it is possible to autonomously mine unlabeled data and form new knowledge, and this is Networking Systems of AI.

Song Liang believes that the future communication network will be the basis of future artificial intelligence, that is, the future network itself will be a distributed intelligence, and the future artificial intelligence system will also exist in the form of a network. "We need a real-time, privately customizable multi-modal network to support application development, superimpose an intermediate layer between the application and the network, adapt the dynamic network to the dynamic application, and allow the artificial intelligence system to evolve from a stand-alone version based on big data The artificial intelligence becomes a distributed artificial intelligence-based network, ultimately realizing a new artificial intelligence paradigm of multi-agent online evolutionary learning."