Evolving and adaptive
Edge Intelligence Systems

AIM and Scope

This special session aims at bringing together academic researchers, industrial professionals, and machine learning practitioners to share their vision and perspectives on the design of evolving Edge Intelligence systems. Edge Intelligence (EI) is becoming popular given the availability of novel Edge Computing platforms that support the execution of AI models directly at the edge of the network without relying on remote entities (i.e., Cloud Computing). Such devices, which usually cost a few tens of euros, can deliver high-performance computing with a low power budget (trillions of operations per watt). However, the design of EI systems requires addressing multiple challenges related to resources utilization, energy footprint, accuracy, model design, deployment, distribution, and many others. For instance, once the system is deployed and put in “production mode”, it requires to deliver the highest performances even if the environmental conditions change, e.g., the luminosity of a scene captured by a smart camera drops. Usually, performance reduction happens since models and applications are built on limited datasets and within controlled environments.

Within this special session, we envision discussing papers that demonstrate how it is possible to efficiently design and deploy Edge Intelligence systems that fully exploit the hardware capabilities of devices and, at the same, they can evolve and adapt to different conditions and contexts. These possibilities might be proved by defining and proposing Edge intelligence architectures, MLOps methodologies, benchmarking of edge intelligence models, design and good practices for edge intelligence systems including extreme environments, methodologies to drive the paradigm shift from intelligent cloud-centric applications to edge intelligence applications, techniques to distribute and orchestrate pipeline workloads across multiple devices. A considerable possibility is offered by soft-computing methods, like evolutionary and bio-inspired techniques, that can efficiently explore the space of possible solutions and find the best trade-off to the given target objective functions, i.e., the size of the model, the energy footprint, the latency, etc. Moreover, we expect the proposition of use cases (e.g., Industrial, agricultural, mining, space, etc.) where these approaches can fully unleash the power of Edge Intelligence.