It's still a legitimate direction to pursue. Once you get to large enough training sets, it's basically the same way our own brains work. We don't perceive or r

It's still a legitimate direction to pursue. Once you get to large enough training sets, it's basically the same way our own brains work. We don't perceive or remember all the details of a building - just "building, style 19B", plus a few extra generic parameters like distance, angle, color and so on. Totally manageable for deep learning to recognize, and perhaps even combine.