eeg
motor imagery classification
a classifier for four motor imagery tasks recorded by eeg: imagined movement of the left hand, right hand, both legs, or the tongue. each trial is 22 electrode channels sampled over 1000 time steps, across nine subjects.
we compared cnns, lstms, grus, and cnn/rnn hybrids. the hybrids won: cnn+gru reached 73% test accuracy, while recurrent models on their own stayed near chance.
sections
1 of 4
the data
four labels, nine subjects, and a signal with a poor signal-to-noise ratio.
eeg measures voltage at the scalp, so each trial mixes the motor signal we want with muscle artifacts, drift, and noise from everything else the brain is doing. the dataset gives 22 channels by 1000 samples per trial and nine labels, of which we classify four.
the structure suggests two different inductive biases. convolution captures spatial relationships between electrodes, and recurrence captures how the signal evolves in time. every architecture we tried is some combination of those two.
arrow keys move between sections.
the models, training notebooks, and the full write-up are on github.
with arvind kalyan and richard jiang, ucla, 2023.