Sparking Light on Deep Learning in EEG Research

Electroencephalography (EEG) is a non-invasive method for recording brain activity, offering valuable insights into neural dynamics with high temporal resolution. Deep Learning (DL), a subset of Artificial Intelligence, is highly effective for analyzing complex, high-dimensional data like EEG by automatically learning features and patterns. This thesis is composed of two parts. The first part investigates DL's application to EEG signal analysis across three tasks: Neural Transcoding, Motor Imagery, and Emotion Recognition, with the goal of improving EEG interpretation and advancing neuroimaging. Here novel DL models are introduced, including the Neural Transcoding Vision Transformer (NT-ViT) which converts EEG signals into fMRI volumes thus connecting these neuroimaging modalities, a multi-stream 1D Convolutional Neural Network for decoding Motor Imagery, and a Subject-Aware Transformer (SATEER) for improving Emotion Recognition accuracy. The second part critically examines limitations in EEG-based DL methods, particularly in Emotion Recognition. Partially drawing on insights from Prof. Giuseppe Placidi, this section highlights challenges like EEG’s inability to capture deeper brain structures involved in emotional processing and the risks of flawed validation protocols, which may inflate model accuracy. These reflections emphasize the need for more biologically grounded methodologies and stronger validation techniques to enhance reliability and generalization. While this thesis introduces new DL approaches for EEG analysis, it also acknowledges the challenges and limitations of current methods. By combining novel techniques with critical evaluation, this work aims to contribute to more accurate and meaningful developments in neuroimaging, cognitive neuroscience, and BCIs, building on the insights of experts.