does the filters of the CNN changes due the back propagation? or they are Do not change?
Hey - not sure if you posted this in the right section of the forums, but then again maybe CNNs are covered in the linalg course as well
Yes, the kernels, or filters, the tiny little windows that move across images and perform convolutions, their values get changed through backpropagation of gradients. The gradients flow through earlier layers (the ones closer to the output of the network), the error gets backpropagated to each of the convolution kernels, and their values are updated in accordance with the gradients.
Just to elaborate on @radek’s response. The values in the filter are actually weights, like in a regular neural net. It’s the same mechanism.