NeuroStream: spectral-spatio-temporal deep learning for visual stimulus classification from EEG

Spectral-spatio-temporal representation In this study, we propose a novel multi-dimensional unified EEG representation designed to capture concurrent spectral, spatial, and temporal dynamics.

ShallowConvNet and DeepConvNet30: The reference shallow and deep convolutional pathways for EEG decoding, representing the filter-bank-like and the deep-hierarchy ends of the design space.

This evaluation aims to identify systematic variations in model performance that may stem from idiosyncratic neural patterns or inherent disparities in stimulus characteristics.

This clustering suggests that the same stimulus triggers highly divergent cognitive responses across different participants, reinforcing the challenge of intersubject variability in EEG decoding.

Another approach we hypothesize to be more effective for EEG decoding is self supervised contrastive loss.