This thesis investigates whether foundation models for EEG can enable reliable zero‑shot and few‑shot calibration in Brain–Computer Interface (BCI) systems. We evaluate REVE, a recently released transformer‑based EEG foundation model, as a universal feature extractor for the P300 paradigm using the BNCI2014‑008 dataset. REVE is kept fully frozen, and subject‑specific classifiers are trained only on its embeddings, allowing us to isolate the quality and universality of the pretrained representations. We extract REVE embeddings for all subjects and train a lightweight MLP classifier under four calibration regimes: 0% (zero‑shot), 5%, 10%, and 50% of each subject’s data. Performance is assessed using Matthews Correlation Coefficient (MCC) to account for class imbalance inherent in P300 datasets. Results show that REVE provides meaningful discriminative structure even without subject‑specific calibration, achieving competitive zero‑shot performance. Increasing calibration data consistently improves MCC, demonstrating that REVE embeddings adapt well to small‑sample personalization. Our findings highlight the potential of EEG foundation models to reduce calibration time in practical BCI systems. By showing that a frozen REVE encoder combined with minimal subject‑specific training can achieve robust performance, this work contributes evidence toward scalable, universal EEG representations and opens pathways for rapid‑deployment BCI applications.
From Foundation Models to Edge Ready CNN Architectures: Distillation, Compression, and Optimization for ERP Decoding
SAFA, NASSER
2025/2026
Abstract
This thesis investigates whether foundation models for EEG can enable reliable zero‑shot and few‑shot calibration in Brain–Computer Interface (BCI) systems. We evaluate REVE, a recently released transformer‑based EEG foundation model, as a universal feature extractor for the P300 paradigm using the BNCI2014‑008 dataset. REVE is kept fully frozen, and subject‑specific classifiers are trained only on its embeddings, allowing us to isolate the quality and universality of the pretrained representations. We extract REVE embeddings for all subjects and train a lightweight MLP classifier under four calibration regimes: 0% (zero‑shot), 5%, 10%, and 50% of each subject’s data. Performance is assessed using Matthews Correlation Coefficient (MCC) to account for class imbalance inherent in P300 datasets. Results show that REVE provides meaningful discriminative structure even without subject‑specific calibration, achieving competitive zero‑shot performance. Increasing calibration data consistently improves MCC, demonstrating that REVE embeddings adapt well to small‑sample personalization. Our findings highlight the potential of EEG foundation models to reduce calibration time in practical BCI systems. By showing that a frozen REVE encoder combined with minimal subject‑specific training can achieve robust performance, this work contributes evidence toward scalable, universal EEG representations and opens pathways for rapid‑deployment BCI applications.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis_final_version.pdf
accesso aperto
Dimensione
4.84 MB
Formato
Adobe PDF
|
4.84 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115848