6G mobile communication networks are envisioned to be AI-native, that is, the provision of AI services to users as well as the network itself would be one of the most essential aspect for its system architecture and design. To provide these services and in order to make correct design decisions, deep understanding of the relationship between the wireless networks and the different AI paradigms along with their various functional aspects is needed. In this paper, we explore the relationship in one such AI paradigm, namely federated learning (FL). More specifically, we aim to look at the problem of hyper-parameter optimization (HPO) for hierarchical federated learning by exploring two design scenarios: (1) HPO is carried out at intermediary aggregators which are implemented at edge servers and (2) HPO is realized by a global server such as a cloud-based approach. Empirical results show that although offloading the HPO algorithm completely to the edge-servers does lower the costs, the hybridized approach whereby the global server realizes the HPO algorithm and the edge-server offers some assistance, can provide at least twice as much savings on training costs. We further discuss the two approaches for providing FL as a native service from the mobile communication network and the implications on the network architecture in light of the discussed design scenarios.

6G mobile communication system is expected to be intelligence inclusive, that is, its system architecture would be designed to provide AI services to anyone, anywhere, including the network system itself. This mandates a clean-slate approach to its design. In this paper, this broad question is addressed from the perspective of providing distributed learning services, more specifically, using the Federated Learning (FL) paradigm. The relationship between two network metrics and the FL performance in a hierarchical federated learning system is explored, in order to determine the new features that the network architecture needs to support. An estimation of the dependency of bandwidth requirements on different ML models is also provided.