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, using Euclidean EEGNet as a baseline, compares it against standard federated learning and centralised training on three motor-imagery datasets spanning diverse channel, subject and class regimes; it observes that personalised SPDNet reaches higher accuracy than both standard federated and centralised training while converging in fewer rounds and communicating fewer parameters than standard federated learning, and that it outperforms every EEGNet configuration on two of the three datasets, although centralised EEGNet outperforms centralised SPDNet.
Interpretation
Personalised federated learning is adapted to the Riemannian SPDNet: all subjects learn a common trunk that builds a latent representation, and each subject keeps its own head. Federated learning for EEG decoding has typically trained a single shared model, which fits each subject poorly because of inter-subject variability; this work brings the shared-trunk plus individual-head personalisation structure into SPDNet, a Riemannian-geometry decoder. The abstract describes the adaptation to SPDNet and states that experiments cover three motor-imagery datasets spanning diverse regimes 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. Accuracy gains and communication efficiency are reported within the same comparison, indicating that personalisation changes not only fit but also convergence and communication cost in federated training. Based on experiments across three motor-imagery datasets; the abstract gives directional conclusions without listing specific accuracy values, round counts or parameter counts.
With Euclidean EEGNet as a baseline, personalised SPDNet outperforms every EEGNet configuration on two of the three datasets, although centralised EEGNet outperforms centralised SPDNet. It provides a cross-comparison of Riemannian and Euclidean decoders under both federated and centralised settings, indicating that the advantage depends on the training setting rather than one method being uniformly better. The abstract explicitly gives the count of two datasets and the reverse result in the centralised comparison, but does not report per-configuration values or statistical tests.
Perspective
The results target motor-imagery EEG decoding and apply to federated settings where one wants to train lightweight decoders without pooling each subject's raw recordings, covering two decoder families, the Riemannian SPDNet and the Euclidean EEGNet, and two comparisons, standard federated learning and centralised training. For readers considering a shared-trunk plus individual-head structure, the work offers a comparison reference across datasets that differ in channels, subjects and classes.
The abstract does not report per-dataset accuracy values, convergence rounds, communicated parameter counts, or whether statistical tests were performed, so the magnitude of 'higher accuracy', 'fewer rounds' and 'fewer parameters' cannot be judged from the available text. The specific names, channel counts, subject counts and class counts of the three datasets are also not listed in the abstract. In addition, the reverse result that centralised EEGNet outperforms centralised SPDNet suggests the advantage depends on the training setting, and its reproducibility on other paradigms or decoders remains an open question.