Transfer of personalized federated learning to Riemannian networks

Determine whether FedPer- and FedRep-style personalized federated learning, which shares a common representation while retaining a client-specific classification head, transfers effectively to Riemannian networks such as SPDNet, and whether Riemannian and Euclidean decoder families benefit from personalization in the same way.

Background

The paper studies personalized federated learning for EEG decoding, comparing a Riemannian SPDNet with the Euclidean EEGNet. Both architectures decompose into a trunk that constructs a representation and a linear classification head. In the personalized protocol, the trunk is aggregated across clients while each client retains its own head.

FedPer and FedRep were originally designed for Euclidean networks. Because EEG data exhibit substantial inter-subject variability and SPDNet operates on covariance matrices on a Riemannian manifold, the paper identifies the transfer of these personalization methods to Riemannian architectures, as well as the relative benefits for Riemannian and Euclidean models, as an unresolved issue. The paper addresses this issue empirically for the studied datasets and configurations, so the unresolved formulation is retained as stated by the authors rather than broadened beyond it.