Our group PS did organize a special session on Embedded AI during the large NL AIC event 2022 (https://nlaic.com/). Our track was very well visited with over 100 people. In this track, Paul Havinga introduced briefly what Embedded AI is, and its major benefits. Subsequently, Jacob Kamminga addressed how biodiversity research can benefit from Embedded AI.
With Embedded AI, (part of) the processing and learning is performed locally, so close to, or even in, the sensor. This opens up a whole new range of applications and reveals many challenges to be solved. The main drivers for this new sector are:
• Greater efficiency and less data transfer: No more transfer of the massive amount of streaming sensor data
• Greater security and privacy: the data stays on the device, privacy-related or confidential data does not leave the system
• Greater responsiveness: as soon as an event is identified, immediate action can be taken, which is essential for, for example, autonomous vehicles or safety systems
• Greater robustness: on-site intervention, without having to rely on network and remote equipment
• Greater functionality: by building AI into things, it can offer new functionalities, so that the machine can, for example, diagnose malfunctions itself
This not only requires a thorough knowledge of the level of the embedded systems and the physical properties of these sensory data, but also requires special algorithms and architectures, taking into account the general context.
In this session, we addressed this from various angles. We looked at how we can help science by monitoring biodiversity by efficiently processing large data streams (speaker Jacob Kamminga, University of Twente), how AI can be embedded in new processor architectures with a fast response time and reduced energy consumption (Menno Lindwer, Grai matter labs), how the high-tech industry uses AI in traffic management sensor systems that guarantee privacy (Michael Dubbeldam, Technolution), and how an SME can add new AI functionality to an existing product (Peter Hoekstra, Datacadabra).
