The classification of EEG during motor imagery (MI) represents a challenging task in neuro-rehabilitation. In 2016, a deep learning (DL) model called EEGNet (based on CNN) and its variants attracted much attention for their ability to reach 80% accuracy in a 4-class MI classification. However, they can poorly explain their output decisions, preventing them from definitely solving questions related to inter-subject variability, generalization, and optimal classification. In this paper, we propose vEEGNet, a new model based on EEGNet, whose objective is now two-fold: it is used to classify MI, but also to reconstruct (and eventually generate) EEG signals. The work is still preliminary, but we are able to show that vEEGNet is able to classify 4 types of MI with performances at the state of the art, and, more interestingly, we found out that the reconstructed signals are consistent with the so-called motor-related cortical potentials, very specific and well-known motorrelated EEG patterns. Thus, jointly training vEEGNet to both classify and reconstruct EEG might lead it, in the future, to decrease the inter-subject performance variability, and also to generate new EEG samples to augment small datasets to improve classification, with a consequent strong impact on neuro-rehabilitation.
