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

By Vernon J. Lawhern, Amelia J. Solon, Nicholas R. Waytowich, Stephen M. Gordon, Chou P. Hung, Brent Lance. Journal of Neural Engineering, 2016

4,485 citations recorded by the source corpus.

Objective. Brain–computer interfaces (BCI) enable direct communication with a computer, using neural activity as the control signal. This neural signal is generally chosen from a variety of well-studied electroencephalogram (EEG) signals. For a given BCI paradigm, feature extractors and classifiers are tailored to the distinct characteristics of its expected EEG control signal, limiting its application to that specific signal. Convolutional neural networks (CNNs), which have been used in computer vision and speech recognition to perform automatic feature extraction and classification, have successfully been applied to EEG-based BCIs; however, they have mainly been applied to single BCI paradigms and thus it remains unclear how these architectures generalize to other paradigms. Here, we ask if we can design a single CNN architecture to accurately classify EEG signals from different BCI paradigms, while simultaneously being as compact as possible. Approach. In this work we introduce EEGNet, a compact convolutional neural network for EEG-based BCIs. We introduce the use of depthwise and separable convolutions to construct an EEG-specific model which encapsulates well-known EEG feature extraction concepts for BCI. We compare EEGNet, both for within-subject and cross-subject classification, to current state-of-the-art approaches across four BCI paradigms: P300 visual-evoked potentials, error-related negativity responses (ERN), movement-related cortical potentials (MRCP), and sensory motor rhythms (SMR). Main results. We show that EEGNet generalizes across paradigms better than, and achieves comparably high performance to, the reference algorithms when only limited training data is available across all tested paradigms. In addition, we demonstrate three different approaches to visualize the contents of a trained EEGNet model to enable interpretation of the learned features. Significance. Our results suggest that EEGNet is robust enough to learn a wide variety of interpretable features over a range of BCI tasks. Our models can be found at: https://github.com/vlawhern/arl-eegmodels.

Subfields: Cognitive Neuroscience

View Paper: https://www.semanticscholar.org/paper/64a2261b178f09e19a4dbbe5296f9f15cee053e3

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

1. Dongrui Wu
2. Xiaorong Gao
3. Cuntai Guan
4. Qingqing Zheng
5. Bingchuan Liu
6. Yonghao Song
7. Yi Ding
8. Qiuhao Zeng

Citations in the corpus

• EEG Based Emotion Recognition: A Tutorial and Review (2022, methodology): “[145] designed an EEG-speciic ConvNet model (EEGNet) by integrating depthwise and separable convolutions.”
• Deep learning for electroencephalogram (EEG) classification tasks: a review (2019, neutral): “1 FC 20 h 2 OUT [94] CNN ELU (conv) SV P300(1) N/A P300(1) P300(1) 3 conv Softmax (FC) All (64/56/22) 1–40 Hz P300(2) 1–40 Hz MI 4–40 Hz 15 93% 1 FC 0.”
• EEG Conformer: Convolutional Transformer for EEG Decoding and Visualization (2022, methodology): “We use exponential linear units (ELUs) as the activation function for nonlinearity following [17].”
• EEG-TCNet: An Accurate Temporal Convolutional Network for Embedded Motor-Imagery Brain–Machine Interfaces (2020, methodology): “Fixed Networks Variable Networks EEGNet*[10] Shallow ConvNet*[9] EEG-TCNET EEGNet EEG-TCNET DFFN [11] Accuracy κ Accuracy κ Accuracy κ Accuracy κ Accuracy κ Accuracy Subject 1 84.”
• Deep learning-based electroencephalography analysis: a systematic review (2019, methodology): “The analysis of model activations was used in multiple studies [93, 97, 120, 172, 186, 214, 227, 231].”
• BENDR: Using Transformers and a Contrastive Self-Supervised Learning Task to Learn From Massive Amounts of EEG Data (2021, background): “While a good deal of evidence [9, 63, 11, 64, 2, 12] supports the advantage of temporally-focused stages (no EEG channel mixing) separate from a stage (or more) that integrates channels, we elected to preserve the 1D convolutions of the original work to minimize any additional confound and to…”
• Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI (2024, methodology): “Some works apply convolutional neural networks (CNN) across and within raw EEG channels to encode spatial and temporal features (Lawhern et al., 2018), while others preprocess the data using short-time Fourier transform (STFT) and employ Graph Neural Network (GNN) on the resulting spectrograms to…”
• SmaAt-UNet: Precipitation Nowcasting using a Small Attention-UNet Architecture (2020, methodology): “In order to reduce the number of parameters without sacrificing a lot of performance, depthwise-separable convolutions (DSC) are used in many recent architectures [26, 27, 28, 16, 29].”
• Physics-Informed Attention Temporal Convolutional Network for EEG-Based Motor Imagery Classification (2023, methodology): “Different DLarchitectureswasproposedforMIclassificationincluding convolutionalneuralnetwork(CNN)[8],[9],[10],[11],[12], [13],recurrentneuralnetwork(RNN)[14],[15],deepbelief network(DBN)[16],autoencoder(AE)[17],andhybridDL models[8],[18].”
• Transfer Learning for EEG-Based Brain–Computer Interfaces: A Review of Progress Made Since 2016 (2020, positive): “Experiments on cross-subject visual evoked potential classification demonstrated that SAN outperformed a support vector machine (SVM) and EEGNet.”
• Federated Transfer Learning for EEG Signal Classification (2020, methodology): “Furthermore, deep learning methods, especially the convolutional neural network (CNN) architectures, have been explored to boost classification performance [18], [19], [20], [21].”
• TSception: Capturing Temporal Dynamics and Spatial Asymmetry From EEG for Emotion Recognition (2021, methodology): “TSception is compared with several deep and non-deep state-of-the-art methods in the BCI domain, namely SVM [33], KNN [34], DeepConvNet [15], ShallowConvNet [15], EEGNet [18], Unsupervised learning [38], FBFgMDM [39], and FBTSC [39].”
• Cross-Modal Epileptic Signal Harmonization: Frequency Domain Mapping Quantization for Pre-training a Unified Neurophysiological Transformer (2025, unknown): “Specifically, we selected ConFormer [54], CSPNet [55], EEGInception [56], EEGNet [57], TinySleepNet [58], and ATCNet [59] as representative baselines for comparison.”