David Wolpert

One can define "fundamental, deep science" in a very literal sense, as the partially hidden parts of science that couple different scientific fields.  

There are many examples of these subterranean parts of the scientific enterprise. They include information theory, computer science theory, and the foundations of mathematical logic. Other examples include unavoidable limitations on how well any machine learning algorithm can perform, and unavoidable thermodynamic costs of running physical systems that perform computation.

David Wolpert works on all of these examples. He is particularly keen to take approaches in one field and apply them to a different field, to gain new insights into that other field.

At the same time, though, he is highly sensitive to how easy it is for scientists to let their mathematical formalizing get ahead of their experimental data. In the theoretical sciences, it is very easy to believe one’s own propaganda, to misconstrue one's castles in the sky for physical reality, to confuse one's arguments for what the universe should be with what the universe actually is.

To combat this tendency in his own work, David tries to exploit real-world data, and in particular real-world engineering data, whenever possible, to correct his theorizing. He is adamant about how important it is "to use the real world to kick me in the ass, and mock my claim to understanding." Accordingly, he has done a lot of applied work designing powerful algorithms for machine learning, for distributed control of aerospace vehicles, and for optimization.