What I have is more likely a theoretical question. Since I am not from a signal processing background it is hard for me to grasp the issue in using convolution in the frequency dimension of a Fourier transformed signal.
What I am trying to do is apply 3D convolution (in a 3D convolution neural network) for a set of Fourier transformed signals recorded by a set of electrodes. The 3 dimensions are the frequency dimension and 2 spatial dimensions of electrodes (x and y).
I understand that what I am trying to achieve is for the model to learn frequency patterns. However, my colleagues feel there's something theoretically wrong in this approach because unlike in time domain, it might not be intuitive to compare among frequencies. I'd value any opinion in this regard or an explanation as to why it is wrong.