Various industries use embedded vision for applications such as automotive, robotics, healthcare and security. Accordingly, the discussion began with the topic of the most promising use cases for embedded vision and the role that artificial intelligence (AI) plays at the edge. The panelists discussed how AI can be used at the edge for various applications such as factory automation, security, quality control, robotic automation and healthcare.

They also discussed how AI at the edge of the network can enable real-time decisions, reduce latency and improve data privacy by processing data locally. What are the biggest challenges when using AI at the edge compared to using it in the data center or in the cloud? Here, the panelists mentioned the limited computing resources and power consumption of edge devices, as well as sometimes the lack of a communication network or the unwillingness of users to rely on a communication network for security reasons. Accordingly, trade-offs between performance, power consumption and cost are essential in embedded vision systems. The panelists emphasized the need for a balanced approach to optimizing the trade-offs: This must include a systematic approach to software development, the treatment of each individual project and the optimization of algorithms, hardware and power management techniques.

Another important issue in the field of embedded vision is ensuring the security and privacy of data. Orr Danon, CEO of Hailo, commented: "The ability to process at the edge of the system is secure, especially when it comes to data privacy. With video processing, you don't really want to know everything in the picture, you just want to get some insights from it. Not collecting everything in one central location is the right way to go, in my opinion." Other panelists pointed out the need for data encryption, secure communication protocols and knowing who owns the data.

The author: Dr. Ekaterina Sirazitdinova is a Senior Data Scientist at Nvidia in Munich.

© Nvidia

Last but not least, the discussion also touched on the question of whether AI training should take place at the edge or in the cloud. The usual approach is to train models in the cloud or on more powerful hardware and then use them on embedded devices for inference - 'cloud-to-edge' or 'server-to-edge'.