3 ms·
After skimming the paper, my first thought was that I want to see how this compares to just adding normal wavelet for Fourier transform (computationally cost
by nabla9 8y ago
After skimming the paper, my first thought was that I want to see how this compares to just adding normal wavelet for Fourier transform (computationally cost and accuracy) steps between NN layers. As I understand it, they are trying to do basically similar thing with new convolution operator.
Wavelet or FFT will not have any extra weights to train and can be highly optimized.