1
$\begingroup$

I am using a 14 channel EEG device. To do away with the need for any handcrafted features, I wish to implement an ML classification task with the EEG data collected using deep neural networks (such as 2D-CNN) directly on the spectrograms.

I get decent results by concatenating all 14 spectrograms (per training example) side by side, to get a single spectrogram. Though this is the only thing that makes sense to me, what is the normal/right way to feed multi-channel spectrograms to a deep neural network?

$\endgroup$
1
  • $\begingroup$ I'd love to see someone on this stackexchange answer your question, but according to this post on meta.stackexchange.com your best bet is probably CrossValidated. Beyond that -- I don't think that there is an established-enough art for there to be a "normal" way to do it, and the "right" way probably depends on the nature of the data and what information you want to extract. $\endgroup$
    – TimWescott
    Feb 3 at 20:54

1 Answer 1

2
$\begingroup$

Concatenation is how this is usually handled, but usually you view your data a as a stack of spectrograms. If you have C channels each with, then for each channel you compute a spectrogram with L time steps, and N bins per timestep, then an established way to view the input as a tensor of shape (N, C, L), this is the convention used for instance in pytorch convolutions.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.