The usual style DL paper that focus on tweaking network architecture is running into a dead end. The improvement is becoming ever more marginal but the cost is

The usual style DL paper that focus on tweaking network architecture is running into a dead end. The improvement is becoming ever more marginal but the cost is skyrocketing. Sadly, bigger and deeper neural networks are previously what major drive force to advance DL research. Meanwhile people turn into RL to approximate non differentiable function, so far it doesn't seem to apply to real world tasks. Give or take 2 to 3 years, if there isn't some major breakthrough, like human level star craft agent, we can then offically announce that this round of DL revolution or renaissance is over.