Is there a standard way to train a convolutional kernel where the kernel has a particular structure, for my specific example, I’d like the kernel to be rotationally symmetric:
c b c b a b c b c
And more generally, I’d like to be able to create a function F parametrized by (a_0, a_1,…) such that the kernel is just F applied to the elements of a matrix of the distances (squared) from the center of the kernel. Then I’d like the parameters (a_0, a_1,…) to be trained directly.
I’ve tried guessing so far but without success.