Depthwise-separable convolution

A depthwise-separable convolution factorizes an ordinary convolution into two cheaper steps: a depthwise convolution that filters each input channel independently, followed by a pointwise (1x1) convolution that mixes channels. This cuts parameters and computation by roughly the ratio of the kernel size to the output-channel count, which is why compact EEG networks (notably EEGNet) adopt it. In neural data the depthwise stage carries a useful interpretation: applied across electrodes after a temporal filter, it learns one spatial filter per frequency band, decoupling the where from the what.

The depthwise multiplier — how many spatial filters are learned per input feature map — is a key capacity knob. Too large and the network overfits the small trial counts typical of BCI; too small and it cannot represent multiple spatial patterns for a single rhythm.