In Search of Explainable and Interpretable Machine Learning with Philosophy and Physics

Machine learning (ML) algorithms are increasingly permeating our everyday lives and public life. They make predictions, but why they decide one way rather than another often remains difficult to understand; in a sense, they are “opaque.” In our project, we aim to understand how this opacity arises and how it might be retroactively resolved. To this end, we intend to interpret the nature of the (implicit) abstractions that ML generates on its own, drawing on insights from physics and other theories of complexity. Our working hypothesis is that the complexity of ML and the difficulty of understanding certain components of the learning process together give rise to the problem of opacity. In this sense, a solution requires not simply “more understanding” or “less complexity,” but a meaningful reduction in complexity. By this we mean adequate abstractions and non-trivial simplifications that ensure a well-founded approach to understanding . In our project, we will develop tools to analyze the complexity of ML algorithms in new ways and identify meaningful reductions from the perspectives of many-body physics and philosophy .