Just because its funny, here is the obligatory reminder that deep convolutional networks are actually implemented as correlation, not convolution.
True, its just the lack of mirroring that differs, which is a linear operation and hence, the network can be considered too have learned the mirrored kernels, but it matters for initialization(bilinear init for conv^T sampling is still incorrectly mirrored in pytorch), and every time someone has put the images of the kernels in a paper its better than even odds they show the correlation, not convolution kernls.
True, its just the lack of mirroring that differs, which is a linear operation and hence, the network can be considered too have learned the mirrored kernels, but it matters for initialization(bilinear init for conv^T sampling is still incorrectly mirrored in pytorch), and every time someone has put the images of the kernels in a paper its better than even odds they show the correlation, not convolution kernls.