Personalised federated learning lets SPDNet EEG decoding beat standard federated and centralised training on three motor-imagery datasets and outperform every EEGNet configuration on two of them
Synopsis
The work adapts personalised federated learning, in which all subjects share a trunk while each subject keeps its own head, to the Riemannian SPDNet, and compares it against standard federated learning and centralised training with the Euclidean EEGNet as a baseline; across three motor-imagery datasets spanning diverse channel, subject and class regimes, personalised SPDNet reaches higher accuracy than both standard federated and centralised training, converges in fewer rounds and communicates fewer parameters than standard federated learning, and outperforms every EEGNet configuration on two of the three datasets, although centralised EEGNet outperforms centralised SPDNet.
Interpretation
The work adapts personalised federated learning to the Riemannian SPDNet: all subjects jointly learn a trunk that builds a latent representation, and each subject keeps its own head for classification. Personalised federated learning has mostly been applied to Euclidean models; here it is extended to the Riemannian SPDNet while EEGNet is retained as a Euclidean baseline, so decoders with two different geometries can be compared within one framework. The summary states explicitly that the adaptation targets SPDNet and that EEGNet serves as the Euclidean baseline; experiments cover three motor-imagery datasets that differ in channels, subjects and classes.
Personalised SPDNet reaches higher accuracy than both standard federated learning and centralised training, while converging in fewer rounds and communicating fewer parameters than standard federated learning. Relative to standard federated learning, personalisation keeps an individual head alongside the shared trunk, easing the poor fit of a single shared model caused by inter-subject variability; relative to centralised training, it achieves higher accuracy without pooling the raw recordings. The conclusion comes from comparative experiments on three motor-imagery datasets, and the summary describes the observations as higher accuracy, fewer rounds and fewer parameters.
Personalised SPDNet outperforms every EEGNet configuration on two of the three datasets, but centralised EEGNet outperforms centralised SPDNet. This comparison places Riemannian and Euclidean decoders relative to each other under both federated and centralised settings, indicating an interaction between geometry and training paradigm rather than uniform dominance of one side. The summary reports the comparison across three datasets and explicitly notes the exception that centralised EEGNet outperforms centralised SPDNet.
Perspective
The result is meant for federated motor-imagery EEG decoding, where recordings from several subjects are not pooled, each subject keeps its own head, and the trunk is shared. It suits researchers and system designers concerned with inter-subject variability, communication cost and privacy constraints, especially teams using light decoders such as the Riemannian SPDNet or the Euclidean EEGNet. The conclusions described in the summary are limited to three motor-imagery datasets and do not state whether other paradigms, other modalities or larger subject populations are covered.
At the summary level, no accuracy values, subject counts, channel counts or class counts are given per dataset, nor are the magnitudes of the reductions in convergence rounds and communicated parameters, so effect sizes cannot be judged. The summary does not explain why personalised SPDNet fails to beat every EEGNet configuration on the third dataset, or why centralised EEGNet outperforms centralised SPDNet. Because only the summary was read here, details of figures, statistical tests and ablations remain open questions.