Real-Time Decision Making with Embedded AI

2021 Smart Transformation: New Innovations

Artificial Intelligence

2021/7/26

On July 22, it was reported that a team led by Professor Su Rijian from the School of Computer and Communication Engineering at Zhengzhou University of Light Industry has made new progress in the project “Research on Non-invasive Temperature Measurement Methods Based on Superparamagnetic Nanoparticles” funded by the National Natural Science Foundation.

This project is the first to apply embedded technology in the field of biological information measurement, achieving significant innovations and breakthroughs in embedded artificial intelligence. The project clarified the relationship between the temperature of the thermal therapy target area and the characteristic parameters of magnetic nanoparticles through embedded AI algorithms, and explored a non-invasive, in vivo temperature field measurement method based on magnetic nanoparticles, which can provide effective temperature and concentration magnetization models as well as biological heat transfer models for temperature measurement in targeted thermal therapy for tumors, significantly improving the treatment effects for malignant tumors.

Real-Time Computing Without Internet Connection

“Embedded artificial intelligence is one of the hottest commercialization technologies in AI today. Embedded AI means that devices do not need to perform large-scale calculations through cloud data centers, but can achieve real-time environmental perception, human-computer interaction, and decision control locally without an internet connection,” said Su Rijian, explaining that embedded AI utilizes embedded terminal devices to perform simplified model training solely through edge computing.

The general view in the industry is that artificial intelligence requires extensive data training through computers to reach knowledge, reasoning, decision-making, and control capabilities similar to or exceeding those of humans. Unlike traditional AI that conducts large-scale data model training in cloud data centers, embedded AI features decentralization, simplified models, reduced training data, and high real-time performance.

Su Rijian stated that 5G technology will give rise to more AI application scenarios, predicting that by 2025, AI will generate a $5.1 trillion application market. Embedded AI can be applied in areas such as healthcare, retail, intelligent transportation, and smart manufacturing.

Reducing Massive Calculations to Small Terminals

According to statistics from the Global System for Mobile Communications Association, the number of connected IoT devices worldwide reached 12.6 billion in 2020. If all devices were to perform cloud computing according to the current AI model training direction, a large portion of IoT terminals would be constrained by network bandwidth speed, unable to achieve high real-time decision-making and control. This would inevitably lead to poor user experiences and even more serious issues, such as high latency in autonomous driving posing driving risks; uploading facial recognition data could lead to privacy information leaks, etc.

“Of course, this does not mean that AI performing model training and decision control in the cloud is contradictory to embedded AI; it is just that the demands of the two are different, and the suitable application scenarios also differ,” Su Rijian explained. The cloud is more suitable for high data throughput and complex model training, while embedded AI has advantages in scenarios requiring high real-time computing.

“It can be said that embedded AI essentially deploys cloud algorithms to terminal devices, which is the marginalization of computation. The biggest challenge is to reduce the massive data and large-scale computations from the cloud to the terminal devices,” Su Rijian said. Specifically, the challenges include: how to increase neural processing units or nodes in chip design based on model training algorithms; how to simplify training models without reducing descriptive accuracy; and how to reduce the amount of training data while ensuring decision accuracy and control precision. “These are not only issues faced by the industry but also technical bottlenecks that our researchers need to overcome.”

END

Source: Science and Technology Daily

Reporter: Ma Aiping

Editor: Lin Xianxian

Related activities of the Digital Expo

Vice Premier Liu He attended the 2021 Digital Expo online

“News Broadcast” focuses on the 2021 Digital Expo

“Focus Interview” focuses on the 2021 Digital Expo

The 2021 Digital Expo concluded successfully

“Digital Salon” enters Guizhou Renhe Data on April 9

“Digital Salon” enters AIRLOOK on April 15

“Digital Salon” enters Xiangshu Technology on April 23

“Digital Salon” enters Guizhou Haohuo on April 27

“Digital Salon” enters AIRLOOK on April 15

“Digital Salon” enters Xiangshu Technology on April 23

“Digital Salon” enters Guizhou Haohuo on April 27

Related posts

Leave a Comment Cancel reply