Bio-signal (EEG) processing for with deep learning at the edge

Brain-Computer Interface (BCI) suffer from high inter-subject variability and limited labelled data, often requiring lengthy calibration phases. This work is part of a broader collaborative effort developed during an internship at Fondazione Bruno Kessler, aimed at building an end-to-end approach that explicitly models subject dependency using lightweight convolutional neu ral networks (CNNs) conditioned on the subject’s identity. The contribu tions presented here focus on the design and implementation of the neural architectures, the hyper-parameter optimization pipeline ensuring robustness and reproducibility, and the interpretability analysis of the learned represen tations. The method integrates two conditioning mechanisms to adapt pre trained models to unseen subjects with minimal calibration data. We bench mark three lightweight architectures on a time-modulated Event-Related Po tential (ERP) classification task, providing interpretable evaluation metrics and explainable visualizations. Results demonstrate improved generalization and data-efficient calibration, highlighting the scalability and practicality of subject-adaptive BCIs.