Linking EEG Channels with Spatial Metadata to Impact the Interpretability and Generalization Ability of Graph Neural Networks in Neurophysiological Applicationsтезисы доклада
Аннотация:We present a novel technique and a specialized machine learning model architecture designed to assist researchers across diverse domains in automatically and explicitly identifying functional patterns in multivariate time series derived from a series of experiments. Our approach formalizes the original task within the realm of machine learning, eliminating the need for domain-specific exploration. The proposed architecture leverages information about the recording device, encoding the initial electrode positions as a graph for transmission to the corresponding model.