Running AI on edge devices poses additional quality of service constraints, as there is typically less centralised technical intervention possible and less ability to scale services up or down. Furthermore, devices are more likely to be intermittently available, or in a dynamic configuration. Consider, for example, a collection of warehouse drones performing order picking. New devices can enter or leave the network at any time. Self-orchestration and task scheduling in such a context will require flexible architectures, potential collaborations between multiple agents. In the context of remote sensing and monitoring applications, an additional challenge can be the need for computational load orchestration, when especially salient events are detected by one sensor that require more compute to analyse than is available on any single device.
One example research initiative at IDLab Ghent is the creation of a distributed network intrusion detection system that can operate in near real-time and aims to classify network traffic into regular traffic and anomalous, potentially dangerous network flows for instance from a cyber-attack.
Another pertinent use case is building a global, shared world view between distributed actors and sensors that are only able to observe their local environment. How do we perform persistent object tracking when objects disappear from local view, only to be picked up by a different sensor at a local time? Research questions like these are addressed by multiple teams in IDLab.
