Research map / Label-efficient and robust learning / Domain adaptation and generalization
SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SPDIM, a source-free, parameter-efficient manifold optimization method that learns one SPD-constrained bias per target domain with information maximization to handle label shift on top of Riemannian alignment for EEG.
Problem
Source-free unsupervised domain adaptation for EEG decoding under conditional and label shifts, where Riemannian Fréchet-mean alignment (RCT+TSM) hurts generalization when class priors differ across domains.
Approach
Derives a generative model with log-linear source covariances and shows Fréchet-mean alignment over-corrects under label shift; then learns a domain-specific SPD-manifold bias (or a scalar geodesic step-size toward identity) on a frozen TSMNet by minimizing an information maximization loss on unlabeled target data.
Key results
- On sleep staging across 426 subjects, TSMNet+SPDIM(bias) reached 69.9% overall balanced accuracy in the patient group vs 64.9% for TSMNet+SPDDSBN, roughly a 5% margin.
- In motor imagery with 0.2 target label ratio, SPDIM(bias) achieved the top cross-subject overall score among the methods reported, with SPDIM(geodesic) at 67.7% and SPDDSBN at 66.4%.
- Simulations confirm RCT compensates conditional shifts only when label ratio is 1.0, and SPDIM outperforms RCT across most configurations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add an SPD-manifold-constrained bias parameter optimized with IM instead of SPDDSBN alignment | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 64.9 -> 69.9 | Helped |
| add an SPD-manifold-constrained bias parameter optimized with IM instead of SPDDSBN alignment | TSMNet · Sleep staging, overall healthy subjects | balanced accuracy: 70.1 -> 73.5 | Helped |
| add the full SPDIM bias parameter rather than the geodesic-constrained step-size parameter (ablation) | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 66.8 -> 69.9 | Helped |
| add the full SPDIM bias parameter rather than the geodesic-constrained step-size parameter (ablation) | TSMNet · Sleep staging, overall healthy subjects | balanced accuracy: 71.4 -> 73.5 | Helped |
| add IM and an SPD-manifold-constrained bias instead of per-domain alignment without IM (ablation) | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 64.9 -> 69.9 | Helped |
| add IM and an SPD-manifold-constrained bias instead of per-domain alignment without IM (ablation) | TSMNet · Sleep staging, overall healthy subjects | balanced accuracy: 70.1 -> 73.5 | Helped |
| use per-domain rather than global alignment, and add SPDIM bias adaptation (ablation) | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 66.3 -> 69.9 | Helped |
| adapt an SPD-manifold bias rather than only the classifier bias (ablation) | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 66.8 -> 69.9 | Helped |
| adapt an SPD-manifold bias rather than classifier parameters (ablation) | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 64.9 -> 69.9 | Helped |
| adapt an SPD-manifold bias rather than the feature extractor and classifier (ablation) | TSMNet · Sleep staging, overall patient subjects | balanced accuracy: 42.5 -> 69.9 | Helped |
| apply RCT+TSM under label shift rather than avoiding marginal alignment | RCT+TSM · EEG classification with label shifts | generalization | Hurt |
| use SPDIM instead of RCT in the generative-model simulations | SPDIM · Simulated binary classification across label ratios | balanced accuracy | Helped |
Surprising findings
- Fitting the Fréchet mean per domain (TSMNet+SPDSBN) hurts generalization under label shift compared with a global alignment.
- Fine-tuning all of the network parameters with the IM loss collapses performance (42.5% overall on sleep staging).
Limitations
- The IM loss can be noisy or outlier-sensitive, sometimes estimating an inappropriate bias that shifts data in the wrong direction and increasing variability.
- No method fully compensates the artificially introduced label shifts in motor imagery experiments.
Experimental setup
Datasets: BNCI2014001 BNCI2015001 Zhou2016 BNCI2014004 CAP Dreem HMC ISRUC
Benchmarks: BNCI2014001 BNCI2015001 Zhou2016 BNCI2014004 CAP Dreem HMC ISRUC
Metrics: balanced accuracy t-test significance
Baselines: RCT Euclidean alignment (EA) STMA SPDDSBN IM (classifier) IM (all) Chambon USleep DeepSleepNet AttnSleep EEGNet EEG-Conformer
Builds on: TSMNet tangent space mapping recentering (RCT) information maximization SPDDSBN
Closest papers by meaning
- SPD domain-specific batch normalization to crack interpretable unsupervised domain adaptation in EEG (NeurIPS 2022 · similarity 0.92)
- Geodesic Optimization for Predictive Shift Adaptation on EEG data (NeurIPS 2024 · similarity 0.90)
- LAtte: Hyperbolic Lorentz Attention for Joint-Subject EEG Classification (NeurIPS 2026 · similarity 0.86)
- NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.84)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.84)
- Riemannian Flow Matching for Brain Connectivity Matrices via Pullback Geometry (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.