Hacker News Reader

Well, you can still probably reduce the training computational cost by feeding the additional FFT features to your NN.

The Fourier transform layer might not be the immediate hidden layer.