I think "giving the rules we already know" is only a good idea if we are 100% sure that we know 100% of the rules necessary with 100% accuracy. The fuzzyness of

I think "giving the rules we already know" is only a good idea if we are 100% sure that we know 100% of the rules necessary with 100% accuracy. The fuzzyness of NNs is what in many cases allows them to work as well as they do, and training on close-to-perfect data is going to preserve this aspect better than close-to-perfect rules could.