In search of explainable and interpretable machine learning with philosophy and physics
Machine learning (ML) algorithms are permeating our everyday and public lives with increasing intensity. They make predictions, but why they ‘decide’ one way and not another often remains unintelligible to us: in a sense they are “opaque”. In our project, we want to understand how this opacity arises, and whether or how it could be retroactively reversed. To do this, we want to interpret the nature of the (implicit) abstractions that ML inherently generates, using 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 this opacity problem. In this sense, a solution does not call for simply “more understanding”, or “less complexity”, but for a sensible reduction of complexity. By this we mean abstractions that are adequate and simplifications that are non-trivial, to ensure access to well-grounded understanding. In our project, we will develop tools to analyze the complexity of ML algorithms in new ways and find reductions that make sense from the perspectives of many-body physics and philosophy.