EEGNet: A Compact Convolutional Network for EEG-based Brain-Computer Interfaces

Objective: Brain-Computer Interface (BCI) technologies enable direct communication between humans and computers by analyzing brain measurements, such as electroencephalography (EEG). BCI processing typically consists of heuristically extracting features for specific tasks, limiting the generalizability of the BCI across tasks. Here, we asked whether we can find a single generalized neural network architecture that can accurately classify EEG signals in different BCI tasks. Approach: In this work we introduce EEGNet, a compact fully convolutional network for EEG-based BCIs. We compare EEGNet to the current state-of-the-art approach across four different BCI classification tasks: P300 visual-evoked potentials, error-related negativity responses (ERN), movement-related cortical potentials (MRCP), and sensory motor rhythms (SMR). We fit 12 different architectures, all with the same number of parameters, to statistically control for the effect of model size versus model performance. Results: We show that one particular architecture performed on average the best over all datasets, suggesting that a generic model can be used for a variety of BCIs. We also show that EEGNet compares favorably to the current best state-of-the-art approach for each dataset across all four datasets. Significance: Our findings suggest that a common simplified architecture, EEGNet, can provide robust performance across many different BCI modalities.