Deep learning for EEG
EEGNet and its relatives, why they are small, what data augmentation does, what foundation models promise, and the honest benchmark result that deep networks often tie the Riemannian baseline on EEG-sized data.
You are skimming: the title, the first figure, and the short version. Switch to Read in the header for the full page, or Deep to open every deep dive.
Deep learning transformed vision and language because those fields have millions of examples. A motor imagery dataset has a few hundred trials per person. That single fact explains most of what you need to know: the networks that work on EEG are tiny, they are designed to bake in what we already know about the signal, and on fair benchmarks they usually tie the best classical methods. They win clearly in two places: when data are large (sleep staging, seizure detection, intracortical speech), and when the features are unknown.
Why EEGNet is shaped the way it is
EEGNetEEGNetA compact convolutional neural network for EEG, published in 2018, that became the standard deep-learning baseline because it is small enough to train on a laptop. Glossary entry (Lawhern et al., 2018) has about two thousand parameters, which is smaller than a single layer of most image networks. Its layers mirror the classical pipeline. First a temporal convolution: a set of learned band-pass filters, one per feature map. Then a depthwise spatial convolution: for each learned filter, a learned spatial filter across channels, which is exactly what CSP produces. Then a separable convolution to combine and summarize over time, and a small classifier. It is CSP plus band-pass plus LDA, made learnable end to end. That is why it works with a few hundred trials and why its filters can be plotted and interpreted.
ShallowConvNet and DeepConvNet (Schirrmeister et al., 2017) are the other standard baselines, similar in spirit. All three are in BraindecodeBraindecodeA PyTorch library of deep-learning models for EEG, including EEGNet and ShallowConvNet, with dataset loaders. Glossary entry with the datasets and training loops built in.
What the benchmarks say
MOABBMOABBMother of All BCI Benchmarks: a library that runs classifiers across many public EEG datasets the same way so results are comparable. Glossary entry runs classifiers across dozens of public datasets with fixed, fair splits. The recurring result: within-subject, the Riemannian tangent-space classifier and EEGNet-class networks are within a few points of each other on motor imagery, P300, and SSVEP, and the network’s small advantage often disappears when hyperparameters are tuned honestly. Cross-subject, both are poor and the network sometimes does a little better because it can be trained on everyone at once. If a paper claims a large deep-learning gain on a small EEG dataset, the first question is the split.
A paper trains a transformer on 200 motor imagery trials and reports 94 percent, versus 72 percent for CSP + LDA. What is the most likely explanation?
Leakage or tuning on the test set. A model with hundreds of thousands of parameters trained on 200 trials cannot generalize by twenty points better than a well-set-up linear method; it can memorize. When results like this are reproduced with honest splits the gap collapses. Project F is where you find this out for yourself.
Where deep learning does win
Large data. Sleep staging has hundreds of thousands of labelled epochs in public datasets; small CNNs reach expert-level agreement and beat feature engineering. Seizure detection on TUH, similarly. Intracortical speech decoding, where a recurrent network maps hundreds of channels of neural features to phonemes, is the most dramatic recent result in the field and is only possible with deep models trained on hours of data from one person.
Unknown features. Where nobody knows what to compute, a network that learns its own filters can find it. The price is interpretability, which for a medical device is not optional.
Transfer. Pretrain on many people or many datasets, fine-tune on the new person with a few minutes of data. This is where EEG “foundation models” aim: large self-supervised models trained on thousands of hours of heterogeneous EEG, fine-tuned per task. Early results are promising for clinical tasks and modest for BCI paradigms. Watch this space; do not assume it has arrived.
Practical rules
Use Braindecode; do not write your own training loop for a first project. Compare every network against the Riemannian baseline on the identical split, in the same script, or the comparison is not real. Augment: time shifts, channel dropout, amplitude scaling, and mixing trials all help small EEG datasets. Report cross-session and cross-subject numbers, because networks overfit within-session more than linear models do. Plot the learned temporal and spatial filters; if they do not look like band-passes over motor cortex, ask what the network learned instead (often eye movements).
Deep dive Self-supervised pretraining for EEG 3 min
Label-free objectives: predict a masked stretch of signal from its context, or tell whether two windows come from the same recording, or reconstruct a window from a corrupted version. Trained on thousands of recordings, a model learns general EEG structure and needs few labels downstream. Reported gains are largest for clinical classification (abnormal versus normal EEG) and for sleep, smaller for motor imagery, which appears to be limited by the signal rather than the model.
Deep dive Interpretability and the FDA 2 min
A device that decides whether a patient is having a seizure needs to be explainable enough that its failures can be anticipated. Regulators have cleared deep-learning seizure and sleep software, with extensive validation and locked models. The regulatory path for a model that keeps learning after deployment is being written now. Phase 5’s regulatory page has the vocabulary.
Explain what this page was about to your roommate in three sentences. No jargon they would not know.