Emerging Technologies for Edge Intelligent Computing Systems

Future next generation networks are expected to outline a wide class of ambitious challenges, mainly concerning the handling and fulfillment of requirements imposed by novel disruptive human-oriented services and applications, for example, the interactive/immersive gaming and the ultimate virtual or augmented reality, involving five senses solicitation. The more and more need of strict requirements in terms of intelligence capabilities, communication latency, and real-time response has witnessed the birth of the new Edge Intelligence (EI) era referring to a set of enabling technologies allowing computation to be performed at the edge of the network, close to the end users, according to the Edge Computing approach, in conjunction with Machine Learning (ML) capabilities. Nowadays, the novel EI paradigm is promoting the development of overarching cross-domains architectural ecosystems for the cost-effective integration of ML into new generation networks, and to contribute to the realization of hyperflexible architectures bringing human-like intelligence into every aspect of the novel EI systems with the aim at properly supporting massive data gathering and processing. Nevertheless, in this context, with the aim at empowering Edge Computing systems with ML modules, we have to take into account that most existing ML algorithms require a large number of data for learning and model training. Meanwhile, clients datasets are typically non-independent identically distributed, and users are often hesitant in sending personal raw data to a third-parties server, representing critical aspects for the practical realization of future intelligent environments. In this picture, the integration of Federated Learning (FL) in several applications scenarios (e.g., anomaly detection, recommendation system, next-word prediction, etc.) is becoming of paramount importance, since it favors users’ privacy preservation enabling a decentralized collaborative training process of a shared global model by using locally collected datasets. The FL approach acts as an iterative data exchange process between end-devices, named clients, and the central server unit, which updates the global model and merges the data processed by the clients side. In this way, personal users data sharing can be exploited, by keeping sensitive information protected.