5G wireless machine learning and communications, a promising combination

 Researchers propose new alternatives to overcome the current barriers that sometimes limit the use of machine learning, which consumes a lot of resources and energy.

Machine learning is radically transforming society, from facial recognition to medical diagnosis, communications, search engines and equipment (drones, autonomous vehicles...). This technology consists of teaching computers one or more tasks from data sets. They become more autonomous and improve their performance as they adapt to the surrounding environment by carrying out operations they were not initially programmed for. Mérouane Debbah, of the Large Networks and Systems Group (LANEAS - Université Paris-Saclay, CentraleSupélec), in collaboration with researchers from the Finnish University of Oulu, recently explored the numerous possibilities that this innovative approach offers. With the arrival of high-performance wireless connections, such as 5G, a new field of applications has been opened up for machine learning.

In most cases, machine learning is organized around a central entity - a server - that collects all the data from peripheral devices through a high-performance wired connection, such as Ethernet. The server is provided with all the data used for learning. However, this organization meets latency and reliability limits when it comes to applications where speed of execution is critical, such as autonomous vehicles.

With 5G wireless connection, researchers are setting a new paradigm: devices themselves are able to access the data they generate. They operate their own models, and share them locally with other devices through this connection, and rarely use the assistance of a central server. Each device enjoys direct and fast access to data required for use, even in the event of an Internet connection failure.

The advantage of this decentralized model is to considerably reduce the latency of the system while improving its reliability. It also improves the management of sensitive personal information, such as medical data, that is not distributed on the Internet but remains stored locally. However, this type of architecture requires sufficient memory and energy capacity on each device to store and process the data collected.

Park, Jihong, et al., "Wireless network intelligence at the edge." Proceedings of the IEEE 107.11 (2019): 2204-2239.