Scientific understanding and artificial neural networks

Abstract:

Artificial neural networks (ANNs) are increasingly used in research in physics and beyond. According to common wisdom, artificial neural networks (ANNs) are good at classification and prediction, but fail when it comes to scientific understanding. One main reason is supposed to be the opacity of such networks. Yet, recently, Emily Sullivan has argued that the opacity of neural networks isn’t the main impediment for understanding a physical system with ANNs. I take this to be an opportunity to clarify how ANNs can, and cannot, contribute to scientific understanding. I start with some examples from theoretical physics and draw on recent philosophical accounts of understanding by Henk de Regt and Alison Hills. I show that ANNs typically do not qualify as vehicles of understanding. Still, they can make a significant contribution to understanding. How much they do depends on what we take to be the most important epistemic subject. Typically, scientists per se do not obtain much understanding if they use an ANN, but a scientist (or a scientific community) plus an ANN as a “coupled system”, as Andy Clark and David Chalmers put it, can be said to understand a bit more.