Tom Sterkenburg: 'Epistemology and theory of machine learning'
The mathematical theory of machine learning offers generalization guarantees that support the epistemological claim that our standard machine learning algorithms are, in fact, good learning algorithms. This is a modern version of the traditional project in the philosophy of science to provide a formal justification for scientific or inductive inferences. But both in philosophy and in machine learning there also exist well-known skeptical results and arguments against the very possibility of inductive justification. This raises the question how a positive story of justification, and in particular a mathematical theory of generalization in machine learning, can exist at all.
In this talk, I answer this question by spelling out the kind of justification that the theory of machine learning offers: general and analytic, yet model-relative. I will argue that this kind of justification fits in a broader epistemological perspective on inquiry. Finally, I will briefly address the recent debate about the apparent failure of classical machine learning theory to explain the generalization of modern algorithms like deep neural networks, and the epistemological contours of a new theory.
If you would like to join this event please email Amit Karmon for the link and further details.
