I think we know some of the basics at least as shown by the results achieved by deep learning(decent computer vision, voice understanding and better natural languege understanding) , and it would be very useful to have optimized hardware to continue research and create products.
Deep learning, as in "series of individually trained autoencoders"? It's a quite specific learning technique. Even if it's currently trendy, there is no particular evidence that any part of the brain works like that. Sure, you can always forget as many details as you need and come back to the basics and say "it's all Hebbian learning anyway, QED", but that's not deep learning anymore.
In one sense it's irrelevant how the brain actually works as long as the tech does what we require. Current methods appear to be good enough to drop them down to silicon.