This computer scientist's simplistic view of the physical sciences is the quest to explain the world from first principles: by deconstructing observed phenomena down to their elementary parts and the laws that govern them, and building useful predictions of large-scale behavior on them.
On the other hand, most modern machine learning and artificial intelligence techniques' astonishing results are based on learning patterns from immense quantities of data, without trying to explain the data through a first-principles model. Often this is by necessity: how could one, for instance, give a truly first-principles explanation of language generation without first having a physically accurate, predictive model of the brain?
In this talk, I will reflect on the differences and similarities of these model-first and data-first approaches and, through examples of recent research, attempt to argue that each holds exciting possibilities for the other.