DeepConvNet & ShallowConvNet

DeepConvNet and ShallowConvNet, introduced by Schirrmeister and colleagues (2017), were among the first convolutional architectures shown to decode raw EEG competitively with filter-bank CSP for motor imagery. ShallowConvNet is deliberately shallow: a temporal convolution, a spatial convolution across all channels, then squaring, average pooling and a logarithm — a pipeline that mimics band-power feature extraction (square gives power, log stabilizes variance), which is why it excels on oscillatory paradigms. DeepConvNet stacks several convolution-pooling blocks with exponential-linear-unit nonlinearities to capture more hierarchical, less hand-designed features.

The paper's lasting contribution was less any single architecture than the recipe that made EEG deep learning work at all: exponential-linear units, batch normalization, dropout, and cropped training (many overlapping time windows per trial as implicit augmentation), together with the demonstration via gradient-based visualization that the networks recovered known sensorimotor-rhythm physiology.