> how do you represent arbitrary spatial relationships on computers in a general and scalable way?

Isn't this essentially what the convolutional layers do in LeNet?