Shouldn't a very complicated perceptron be capable of addition if the problem is extracted from an image? Isn't that what the individual neurons do?