Good morning,

I am coming from learning machine learning convolution for neural nets and was wondering about cross-correlation vs convolution.

I referenced this answer here: What's the difference between convolution and crosscorrelation?

But I fail to understand the practical difference that a mirrored 'filter' (not sure if that is the correct term in this context) produces when using convolution rather than cross-correlation. It seems that either method contains different representations of the same data. Whether (as in the link above) it is X+Y or Y-X, they both contain similar, albeit opposite, data.

Is this simply used to adjust the direction of the vector, as seemingly the magnitude would remain unchanged? Or am I missing some subtleties?

Thank you!

  • $\begingroup$ Please explicate why I get a downvote to help me make better questions in the future =) $\endgroup$ – Free Url Dec 17 '17 at 14:45
  • $\begingroup$ Downvote button says to use when the question does not show any research effort, or if the question is unclear or not useful. Is my research reference unsatisfactory, or is there some ambiguity in my question that I could change... or is someone just having a bad day? =( ... (<- turn that frown, upside down!) $\endgroup$ – Free Url Dec 17 '17 at 14:49
  • $\begingroup$ I have upvoted in order to compensate this downvote which - I don't hesitate to say - is ridiculous, surely done by somebody that hasn't the faintest idea about these issues. $\endgroup$ – Jean Marie Dec 17 '17 at 15:07
  • $\begingroup$ Thanks @JeanMarie, your validation that I am not doing something blatantly unacceptable is much appreciated! People are Quick Draw McGraws when it comes to downvoting in several exchanges it seems. Much appreciation! $\endgroup$ – Free Url Dec 17 '17 at 15:08

From more research I performed, per a Coursera course in which I am enrolled (DeepLearning.Ai), in "signal processing or in certain branches of mathematics, doing the flipping in the definition of convolution causes the convolution operator to enjoy [the associativity property]" - Professor Andrew Ng

The 'flipping' referring to using convolution rather than cross-validation, both terms used in the mathematical context.


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.