Personalised federated learning for Riemannian and Euclidean EEG decoding

• FL fits subject-by-subject EEG collection by exchanging model parameters rather than pooling recordings that confidentiality rules keep from being pooled.Source: §1 · p. 1
• For SPDNet, personalised FL outperforms both standard FL and centralised training on all three datasets.Source: §1 · p. 1
• Whether FedPer [13] and FedRep —designed for Euclidean networks—transfer to SPDNet and benefit both decoder families similarly remains open.Source: §1 · p. 1
• For EEGNet, personalised FL improves on standard FL on two of three datasets but remains below centralised training and brings no speed-up.Source: §1 · p. 2

2.1. EEG decoding architectures

• EEGNet applies temporal, depthwise spatial, and separable convolutions to raw trials before a linear class-scoring layer.Its blocks also use batch normalisation, ELU activation, average pooling, and dropout.Source: §2.1 · p. 2
• SPDNet processes spatial covariance matrices through BiRe blocks that reduce dimension while preserving positive definiteness.BiMap weights lie on Stiefel manifolds, and ReEig raises eigenvalues below ε to ε before LogEig and softmax classification.Source: §2.1 · p. 2
• Both decoders divide into a representation-building trunk and a linear classification head, but SPDNet trunks lie on product Stiefel manifolds whereas EEGNet trunks are Euclidean.The SPDNet head is a softmax layer; EEGNet's is its final linear layer.Source: §2.1 · p. 2

2.2. Federated protocols

• Standard FL transmits and aggregates every learnable parameter, producing one shared model despite multisubject differences in spatial patterns, signal quality, and discriminative rhythms.For SPDNet, arithmetic averaging cannot directly aggregate Stiefel BiMap weights, so ProjAvg [9] projects their mean back onto the manifold layerwise.Source: §2.2 · p. 2
• Personalised FL exchanges only the trunk, while each client trains and retains its own head throughout each round.The server aggregates EEGNet trunks with FedAvg and SPDNet BiMap weights with ProjAvg.Source: §2.2 · p. 2

3.1. Compared configurations

• SPDNet processes covariance matrices through BiRe blocks into a linear space, while EEGNet uses temporal, spatial, and separable convolutions before its linear head.Both architectures are split into a trunk and a linear head.Source: Figure 1 · p. 3
• The two federated protocols differ in whether the head is aggregated or kept local, isolating head personalisation over a common trunk; pooled training serves as the reference.Table 1 identifies the federated configurations and defines pooled training as one model trained on all clients’ data.Source: §3.1 · p. 3, Table 1 · p. 3

3.2. Setup

• The three MOABB motor-imagery datasets were selected to vary in channel count, number of subjects, and number of classes; signals were filtered to 8–32 Hz and resampled to 128 Hz.Source: §3.2 · p. 3
• SPDNet uses one BiRe block, while EEGNet uses its unchanged architecture; the BiMap dimension and batch size are selected on centralised runs and held fixed elsewhere.The rectification threshold is set from the input-covariance eigenvalues of the first three subjects and shared across configurations.Source: §3.2 · p. 3
• Each subject’s trials are split per seed into class-stratified training, validation, and test sets in 75/10/15 proportions, and reported results average ten seeds.Training uses one local epoch per client; validation loss controls learning-rate reductions and stopping, with a 3000-epoch-or-round budget.Source: §3.2 · p. 3

Effect of personalisation

• Standard FL falls below centralised training for both architectures; the reported SPDNet drop is 1.4–3.8 points, while EEGNet drops 17.8 and 13.6 points on BNCI2014-001 and Weibo2014.On Cho2017, EEGNet personalisation brings no gain over standard FL, and both federated configurations remain below centralised training.Source: §4 · pp. 3–4
• SPDNet personalised FL gains 3.4–15.7 points over standard FL and exceeds centralised training on all three datasets, with the Cho2017 gap within one standard deviation.Table 3 reports test accuracies; the prose qualifies the Cho2017 comparison.Source: Table 3 · p. 4, §4 · p. 4
• The authors suggest local heads may partly compensate for inter-subject shifts while a common trunk learns representations from all subjects.They note that each client is one subject and EEG signals vary strongly between subjects.Source: §4 · p. 4

Riemannian versus Euclidean decoding

• Centralised EEGNet outperforms centralised SPDNet by 2.3–7.3 points, but personalised SPDNet leads every configuration on BNCI2014-001 and Weibo2014; centralised EEGNet remains ahead on Cho2017.On the first two datasets, personalised SPDNet scores 68.3 and 56.8 against centralised EEGNet’s 63.7 and 52.0; on Cho2017, the comparison is 65.9 versus 67.9.Source: §4 · p. 4

Convergence and communication

• SPDNet personalised FL reaches within one point of its reported accuracy in 150–270 rounds, versus 600–810 for standard FL, and cuts parameters sent per round by 61%, 55%, and 21% across the datasets.For EEGNet, both federated configurations reach within one point in 90–200 rounds, so personalisation brings no speed-up.Source: §4 · p. 4

5. CONCLUSION

• Across the three motor-imagery datasets considered, personalised federated SPDNet is more accurate than centralised training and faster and cheaper to train than standard FL.The authors conclude that a common Riemannian representation with subject-specific classifiers appears well suited to inter-subject variability in federated learning.Source: §5 · p. 4
• For a new subject with unlabelled trials, the head would need to be built without labels; cross-subject domain adaptation is offered as one possible direction.Source: §5 · p. 4
• Although trials stay on site, transmitted parameters may still leak information; the authors identify differential privacy of the trunk as a future direction.Because heads remain on clients, only the trunk would need to be privatised.Source: §5 · p. 4

6. COMPLIANCE WITH ETHICAL STANDARDS

• The study used open-access human-subject data from the MOABB benchmark; ethical approval was not required.Source: §6 · p. 5