Cross-Session BCI Transfer: A Matched Comparison of Global and Selective Pooling
Date:
Authors: Yiming Shen, David Degras (University of Massachusetts Boston)
Summary
This poster presents a matched comparison of three source-pooling strategies for cross-session EEG transfer in Brain-Computer Interfaces: Global Source Pooling (GSP), Minimum Distance Pooling (MDP), and Bridge Domain Pooling (BDP).
Key Findings
On the BNCI2014_004 dataset:
- GSP gives the highest matched final accuracy (mean 0.7248), winning 22/24 matched comparisons against both MDP and BDP.
- MDP retains target-relevant source sessions via CI-Gating, showing higher within-pipeline DA lift (+1.72% vs +1.33% for GSP; BH-adjusted p = 0.0129).
- BDP splits sources into explicit Bridge/Far roles for robust transfer under severe shift, achieving the largest single DA lift (+4.90% under CSP + Elastic Net + SA).
- Subject dominates overall variance (raw $\eta^2$ = 93.88%); among controllable factors, DA choice explains ~10.8x more variance than pooling identity.
Pooling Strategies
| Strategy | Main Idea | Primary Strength |
|---|---|---|
| GSP | Merge all sessions | Final accuracy baseline |
| MDP | Retain target-relevant sessions | Higher DA lift under source heterogeneity |
| BDP | Split into Bridge/Far subsets | Explicit source roles for severe drift |
Conclusion
Where cross-session drift is relatively mild, GSP is the strongest accuracy baseline. MDP and BDP become better motivated for more heterogeneous or severely shifted targets.
References
- Barachant et al. Multiclass BCI classification by Riemannian geometry. IEEE TBME, 2011.
- Jayaram et al. Transfer learning in brain-computer interfaces. IEEE CIM, 2016.
- Shenoy et al. Towards adaptive classification for BCI. J Neural Eng, 2006.
- Wang et al. Characterizing and avoiding negative transfer. CVPR, 2019.
- Zanini et al. Transfer learning: A Riemannian geometry framework with applications to BCIs. IEEE TBME, 2017.