EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces

EEGNet is a compact convolutional neural network designed for EEG-based brain–computer interfaces. Rather than building bespoke feature extractors and classifiers for each BCI paradigm, the authors use depthwise and separable convolutions to construct an EEG-specific model that encapsulates well-known EEG feature-extraction concepts, aiming for a single architecture that stays as compact as possible while working across paradigms.

Evaluated within-subject and cross-subject across four paradigms—P300 visual-evoked potentials, error-related negativity, movement-related cortical potentials, and sensory motor rhythms—EEGNet generalized better than reference algorithms and matched their high performance even when only limited training data was available. The authors also demonstrated three approaches to visualize a trained model's learned features, showing the network learns robust, interpretable features and making it broadly usable for BCI research.