Deep Learning

i was kind of surprised by the set up. Is basically like, imagine, you a simple case. And you train it separately to do the same exact task on set one of imputs and outputs and set two. This is not really about an idea for making deep learning work better,. But it's an intuition i had that i wasn't really able to find evidence for. I mean, what things seemed obvious, but just felt like this should work, who really should work, but it just doesn't work.