I'd love to see this network tested on hybrid breeds (but still trained on pure breeds) to see if the probabilistic output would be divided between the two parent breeds.