Conference on Cognitive Computational Neuroscience

Artificial Neural Networks (ANN) are becoming the standard machine learning method in a variety of scientific fields. They are currently being increasingly investigated for decoding and modeling brain signals such as electroencephalography (EGG) and functional magnetic resonance imaging (fMRI). However, ANNs are relatively underexploited for Magnetoencephalography (MEG) data classification and modeling. Here we describe a simple and intuitive convolutional neural network (CNN) architecture for MEG and test its performance on three MEG decoding tasks (n=643 subjects). Our model, specifically built with interpretability and speed in mind, achieved 99\% decoding accuracy on the 643 subject identification task, 76.9% on sex prediction and 99.4% for auditory vs visual stimulus classification). In addition to comparison with benchmark models, we showcase tools for latent space visualizations that help uncover the underlying brain mechanisms used by the network. Our work contributes to ongoing efforts within the MEG/EEG community to enhance the usability and interpretability of DL-driven brain decoding.