Association between Generalization Performance and Interpretability in Deep Learning for Cognitive Load Recognition with Frontal EEG

Abstract

Cognitive load recognition with electroencephalography (EEG) by deep neural network (DNN) has been gathering attention for well-being applications in daily life; However, generalization performance for untrained data has not been investigated thoroughly. We hypothesized that generalization performance may depend on psychophysiological plausibility, and evaluated the performance using interpretable DNNs. The results suggested that a model with highly salient features throughout the EEG frequency bands associated with cognitive load have high generalization performance.