Cell-Free Massive MIMO: A Suitable Architecture for Federated Learning

Machine learning (ML) has become essential for processing data collected by wireless devices, for example, to carry out inference and predictive tasks. The training of such ML models could be done at an edge server, but that would require wireless transfer of vast amounts of data from the devices, leading to large communication-related costs as well as data privacy concerns. Federated Learning (FL) is a promising alternative that can greatly reduce the communication costs by performing local computations on the devices and then sharing model update parameters with the server in an iterative manner. The wireless traffic related to FL has two main characteristics: many devices must transfer equally many bits in the uplink and receive a common update message in the downlink. In this talk, we will show how the physical layer of Cell-free Massive MIMO is ideal for supporting FL. We will describe recent research results showing how the uplink and downlink can be optimized to achieve a preferable tradeoff between energy consumption and delay.