The accurate detection of motor imagery (MI) from electroencephalography (EEG) is a fundamental, as well as challenging, task to provide reliable control of robotic devices to support people suffering from neuro-motor impairments, e.g., in brain-computer interface (BCI) applications. Recently, deep learning approaches have been able to extract subject-independent features from EEG, to cope with its poor SNR and high intra-subject and cross-subject variability. In this paper, we first present a review of the most recent studies using deep learning for MI classification, with particular attention to their cross-subject performance. Second, we propose DynamicNet, a Python-based tool for quick and flexible implementations of deep learning models based on convolutional neural networks. We showcase the potentiality of DynamicNet by implementing EEGNet, a well-established architecture for effective EEG classification. Finally, we compare its performance with the filter bank common spatial pattern (FBCSP) in a 4-class MI task (data from a public dataset). To infer cross-subject classification performance, we applied three different cross-validation schemes. From our results, we show that EEGNet implemented with DynamicNet outperforms FBCSP by about 25%, with a statistically significant difference when cross-subject validation schemes are applied. We conclude that deep learning approaches might be particularly helpful to provide higher cross-subject classification performance in multi-class MI classification scenarios. In the future, it is expected to improve DynamicNet to implement new architectures to further investigate cross-subject classification of MI tasks in real-world scenarios.
