5 ms·
It's actually cross-correlation, not convolution? (The two are only equal if kernel is symmetric)
by krackers 3y ago
It's actually cross-correlation, not convolution? (The two are only equal if kernel is symmetric)
- esalman 3y agoYes. Convolution itself has a lot of very useful properties in signal processing (transformation between time and frequency domain for easier compute of LTI systems etc.) But those are lost from the ML literature.
- sorenjan 3y agoIt's just a matter of mirroring the kernel, and since the kernels are learned I don't really see the big deal. I guess the usage of the term convolution comes from image processing and other DSP where it used to be actual convolution, and now the operation is basically the same even if you might not even care about the actual values in each kernel.