Reboning AIOT

21st Century Business Herald reporter Bai Yang Beijing report

On July 15th, on the opening day of 2022 Viewing Technology, the founder and CEO Yinqi of the United Nations proposed the latest definition of AIOT, that is, "AIOT = AI+IoT+Space".

Yinqi said that since its establishment, AIOT has been a strategic keyword that has visited the core. Among them, "AI" and "IoT" complement each other. AI is an evolutionary algorithm capabilities, and IoT is a device carrier combined with soft and hard.

On this basis, the concept of "space" emphasizes the concept of "space". Yinqi said that space is a closed loop of application scenarios. In the past two or three decades, the continuous evolution of the Internet, 5G, AR, VR and other technologies has brought tremendous changes to the virtual world. At the same time, the transformation of technology in the physical world has not changed fundamental changes. Therefore, in the future, making the physical world better will be the direction of AIOT practitioners.

According to Yinqi, the current view of the AIOT strategy has built a "2+1" technical research system. Among them, "2" refers to the two major AI technology systems of "basic model scientific research" and "scale algorithm production", and "1" refers to the IoT technology system with calculation of photography, including AI sensors and AI robots.

It is reported that this complete scientific research system covers a full link from basic research, algorithm production to hard and hard -and -hard integrated products. Indian said that the "2+1" AIOT core technology scientific research system will be an important cornerstone to support the future success of AIOT's commercialization.

Among them, in response to the basic models of AI capabilities, it is proposed that "big" and "unity" have become the latest trend of visual AI basic research.

Zhang Xiangyu, the person in charge of the basic scientific research of the Institute of Everbright, said that "big" mainly refers to the large model of AI, that is, the use of big data, large computing power, and large ginseng quantity to improve the expression of models, so that the AI ​​model can be suitable for multiple tasks. A variety of data and various application scenarios.

However, Zhang Xiangyu also proposed that although "big" is one of the important shortcuts to improve the performance of the AI ​​system, it does not mean that the large number of ginseng quantity, large calculation, and large data are one -sided. The powerful model will generate greater computing expenses, making the overall income very limited.

Therefore, from perspective, "big" should be driven by innovative algorithms, giving full play to the power of big data and large computing power, and expanding the boundary of AI cognition.

As for the "unity", a series of algorithms derived from different research fields of AI vision in different research fields were on the bottom layer.

Zhang Xiangyu said that through unified algorithms and models to represent various data and tasks and tasks, it will produce a simpler, powerful and general system.

According to Zhang Xiangyu, under the concepts of "big" and "unified", the scientific research of the basic models of the basic model focuses on the four directions of general images of large models, video understanding, computing large models and autonomous driving perception models, and has already been perceived the big model, and it has already already been. A number of scientific research results have been achieved.

In addition, AIS (AI Service), a self -developed algorithm production platform was released on the same day. It is reported that AIS is based on the BRAIN ++ system and constructed a set of productivity tool platform covering the full -linking link. The launch of AIS makes the mass production of algorithms really possible.

At present, the AIS platform has been able to support more than 100 business model training, which can be completed in 2 hours as soon as possible. Practice data shows that algorithm R & D personnel use BRAIN ++ and AIS platforms to achieve an average acceleration of intelligent labeling 30 times and automatic learning training acceleration by 4 to 20 times.

- END -