Why to use odd-sized convolution kernels

The most basic concept in an digital image is a pixel value. The idea of convolution is to capture information from a neighborhood of pixels around an anchor pixel. In typical convolution without stride, we apply convolution filters on each pixel (that then serves as the anchor pixel).

To ensure a symmetric filter we have to pick an odd kernel size, since this requires a side length of $2*n+1$.

If you use an asymmetric filter instead (which is possible, of course), you'll introduce aliasing errors because you are effectively centering your kernel on a sub-pixel level.

Great explanation can be found here