Abstract:

In this talk I will discuss two models of how science progresses and their implications on AI research. Science may progress in leaps, where much effort needs to be spent on hopping from one discovery to another. If many intermediate steps between discoveries are available, however, then scientific progress is akin to climbing a ladder. Powerful search methods are necessary for making leaps, while climbing ladders requires the necessary intermediate steps to be available. I will illustrate these points by drawing on insights from my own recent work, the less recent history of deep learning, and by drawing an analogy with natural development. I will conclude by speculating on the future of scientists given the recent results on Navier-Stokes.