Personalised federated learning for Riemannian and Euclidean EEG decoding
Thibault Pautrel, Florent Bouchard, Ammar Mian, Guillaume Ginolhac
stat.ML · cs.LG
Quick verdict
Reports head-personalised FL lifting federated SPDNet over both standard FL and centralised training on three EEG datasets, though one centralised win lacks statistical support.
Summary
This paper adapts FedPer-style personalised federated learning (FL) to the Riemannian SPDNet for EEG decoding: clients share a common trunk aggregated on the manifold with ProjAvg, while each subject keeps a private linear head. Across three public MOABB motor-imagery datasets (BNCI2014-001, Weibo2014, Cho2017; 9-52 subjects as clients), it compares standard FL, personalised FL and centralised training for SPDNet and EEGNet, averaging about ten random seeds. Reported results: personalised SPDNet beats standard FL on all three datasets (gains of 3.4 to 15.7 points) and exceeds centralised SPDNet on all three, with the smallest gap (Cho2017, 2.0 points) within one standard deviation; it also converges in fewer rounds and transmits fewer parameters per round. For EEGNet, personalisation helps on two datasets but never reaches centralised performance, and centralised EEGNet stays ahead on Cho2017. Limitations: no comparison with FedRep or with the subject-wise Riemannian recentering that motivates the work, no locally-trained baseline, no statistical tests, and no code-availability statement.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Personalised FL with a local head makes federated SPDNet more accurate than both standard FL and centralised training on all three motor-imagery datasets.
- Personalised SPDNet converges within one point of its final accuracy in 150-270 rounds versus 600-810 for standard FL, and transmits only the trunk.
- Federation reverses the centralised ranking between architectures on two of three datasets: centralised EEGNet beats centralised SPDNet everywhere, but personalised SPDNet is best on BNCI2014-001 and Weibo2014.
- For EEGNet, personalisation improves over standard FL on two datasets but never reaches centralised performance and brings no convergence speed-up.
Strengths
Fair, tightly controlled comparison
Standard FL and personalised FL differ only in whether the head is aggregated (Table 1), isolating the personalisation effect; three MOABB datasets (Table 2) span 22-64 channels, 9-52 clients and 2-7 classes, and every number is a mean over ten seeds redrawing split and initialisation.
Honest reporting of negative evidence
The paper explicitly reports that centralised EEGNet beats centralised SPDNet everywhere and that personalisation leaves EEGNet below centralised training, rather than burying inconvenient baselines.
Practically meaningful gains for the Riemannian model
Communication analysis is quantified
Weaknesses
No inferential statistics
Missing baselines, including one cited as motivation
Possible tuning asymmetry across configurations
The communication advantage is largely definitional
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This ICASSP-style short paper asks whether head personalisation fixes federated EEG decoding, and whether the answer differs between a Riemannian decoder (SPDNet) and a Euclidean one (EEGNet). Building on the authors' federated SPDNet (FedSPDNet, [9]), it applies FedPer [13]: a shared trunk plus a private linear head, with BiMap weights aggregated by the ProjAvg polar retraction (Eq. 1). Three configurations (Table 1) - centralised, standard FL, personalised FL - are evaluated on three MOABB motor-imagery datasets (Table 2) chosen to span channels (22/60/64), clients (9/10/52) and classes (4/7/2). Central claims: (i) personalised SPDNet beats standard FL by 3.4-15.7 points and beats centralised SPDNet on all three datasets; (ii) it converges in 150-270 rounds versus 600-810 for standard FL and transmits 21-61% fewer parameters per round; (iii) federation reverses the architecture ranking on two of three datasets, while centralised EEGNet remains the single best model on Cho2017.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does personalised federated learning beat centralised training for EEG decoding?
For the Riemannian SPDNet decoder with a private per-subject classifier head, the paper reports higher test accuracy than both standard federated learning and centralised training on all three motor-imagery datasets studied, though the smallest margin (about 2 points on Cho2017) falls within one standard deviation across seeds. For the Euclidean EEGNet, personalisation helps against standard FL but never reaches centralised performance.
What stays local in personalised federated EEG learning?
Each client shares a common feature extractor (trunk) that is averaged on the server every round, while its classifier head stays on the device and is never transmitted. Only trunk parameters cross the network, which cuts per-round communication by roughly the trunk's share of all parameters (21-61% in the paper's models).
Why does standard federated learning underperform on multi-subject EEG?
Because EEG signals vary strongly between subjects, a single shared model - one decision rule for everyone - fits none of the subjects well. Keeping a subject-specific head lets each client adapt its decision rule while the shared trunk still benefits from all subjects' data.
What is SPDNet and how does it differ from EEGNet?
EEGNet is a compact convolutional network applied to raw EEG trials. SPDNet instead takes the trial's spatial covariance matrix - a symmetric positive definite matrix - and processes it through layers that preserve that geometry (BiMap projections onto the Stiefel manifold, eigenvalue rectification, matrix logarithm) before a linear classifier. The paper reports SPDNet degrades less than EEGNet under federation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 2 learning curves cannot be analysed from extracted text; convergence claims rely on the round ranges stated in Section 4.
- Fig. 1 architecture diagram and Table 2/3 formatting were extracted as plain text; only values quoted in prose could be cross-checked.
- arXiv listing states 24 Sep 2026 while reference [9] is 'Accepted for publication' at EUSIPCO 2026; extraction appears complete, so confidence remains high.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.