I am learning about the multi-modal deep learning models and the papers I am reading are very confusing on one point in the process: "feature extraction" and "data normalization" seem to be used interchangeably.

I always thought that feature extraction was the way to select certain attributes of the data instead of normalizing / scaling it. Am I wrong?

  • 1
    $\begingroup$ Can you provide citations to some of these papers? Can you quote some passages where the terms are used synonymously? (Unfortunately, terms are used in widely varying & inconsistent ways.) $\endgroup$ – gung - Reinstate Monica Aug 12 '15 at 20:27
  • $\begingroup$ normalization is moving the numeric values to avoid roundoff, center your coordinates in the center of the clouds, and account for the gross features with gross models so the high-tech tools spend all their horsepower figuring out the hard stuff. Feature extraction - that is the hard stuff. normalization of them is getting the rocket to Cape Canaveral but feature extraction is getting the rocket into orbit. Stone age man can get to the island, but it takes 20th century technology to get into orbit. $\endgroup$ – EngrStudent - Reinstate Monica Aug 12 '15 at 21:26

Caveat: I am not an expert, and my ONLY experience with these terms is from the CUDA webinars.

In the CUDA GPU library webinars, they used data normalization for adjusting things at the pixel level, where each value is mapped to a different value.

Feature extraction would be identifying lines and color regions, mapping several pixels to one feature, like a red blob, or a line.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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