Why does a CNN reuse the same small filter at every position of the input (weight sharing), instead of learning a separate weight for every pixel?