PaperKey Ideas
[2015] Convolutional neural network for multi-category rapid serial visual presentation BCI
  • amplitude and latency of P300 have large variance between subjects and within subject
  • categorise 5 classes (planes, faces, cars, eggs, watches), 360x360px images presented at 10 Hz
  • remove 3 subjects (out of 5) due to blinking, technical, and uncomfortable
  • architecture:
    • input: 64 channels, 64 time samples
    • 1st convolutional layer, spatial convolution, using filters of size 64×1, learning features which represent a spatial distribution across the scalp
    • max-pool to reduce dimension, size 3, stride 2
    • then temporal convolutional layer, find temporal patterns in the signal that represent the change in amplitude of the spatial maps learned in the first layer
    • 2 fully connected, 2048 and 4096
  • spatio temporal regularisation
    • first convolu learn spatial, that change slowly in time by adding penalty to the cost function
    • encourage small differences between consecutive temporal values
[2011] Convolutional Neural Networks for P300 Detection with Application to Brain-Computer Interfaces
  • challenges: user might not response (focus on the flashing light) at the right moment
  • P300 wave is an event related potential (ERP), corresponds to positive deflection in voltage at 300ms latency
  • classification 2 class, corresponds to P300 wave or not
  • normalised
    • size divide by 2, signal sampled at 120 Hz
    • bandpass filtering at 1-20 Hz
    • related work did this: remove eye movement by independent component analysis
  • 5 layers
    • convolu by space
    • convolu by time
    • fully connected
  • instead of using 64 electrodes (channels), they use 8
  • single classifier
  • multiclassifier
  • use first layer's weight to check feature importance, if close to 0 means
  • discriminant power is low