| HN Mirror

Maybe I have to elaborate a bit here. Neural Network sounds fancier than it really is. The steady state visually evoked potentials work as I recall in the way that when a person stares at a light source which blinks in a particular frequency then one can measure an EEG signal over the visual cortex which has the same frequency and multiple of it, as the light source. For example if a person concentrates on a blinking LED with 4Hz then a signal with dominant 4Hz and 8Hz, 12Hz and so on can be measured over the visual cortex. I think in this explanation are some details missing ;-) We used LED sources with different frequencies which represented different actions like left, right, up, down etc. The Neural Network was trained with the frequency spectrum as input and not with the raw EEG signal. However I think the correct and easier solution was to use a filter bank (comb filter) to determine the frequency of the currently measured EEG signal. But back then I read the first time about Perceptrons, XOR problem and Neural Networks. It was cool and I wanted to try it out. The challenging part back then 2010 was that there was no TensorFlow or Pytorch. I was equipped with Matlab, the Signal Processing Toolbox and my dangerous superficial knowledge about Neural Networks. I implemented backpropagation and gradient descent myself for fun and the output was a matrix which represented the weights of the Neural Network. The "inference engine" was a matrix multiplication. The matrix with the weights multiplied with an input vector and as the activation function I used tanh(). Surprisingly the classification worked very well and it was fun to implement all that stuff. This all sounds like grampa talking about WW2 again :-)