80-326 Epistemology of Machine Learning
9 units, Carnegie Mellon University
Learning is hard Programming a computer is also hard. Better that computers learn on their own from data how best to serve us. That is the goal of machine learning ML, which is arguably the most successful branch of artificial intelligence. The very idea raises some natural, fundamental questions. What, exactly, is the goal of learning Is it maintenance of consistency among our beliefs Bayesian statistics, or is it a matter of estimating or predicting quantities in nature frequentist statistics Is the goal prediction or truth control or understanding actual fact or necessary causal relations Are predictions expected to be reliable beyond the narrow circumstances of training Can learning be guaranteed to succeed by a specified time Does success entail some detectable mark or sign of success that can be used to terminate the learning procedure What roles do causality and simplicity play in learning, and how Those questions arise spontaneously for the reflective ML researcher, but they also cross over into the foundations of statistics, the philosophy of science, and epistemology, the traditional philosophical study of the nature of knowledge and justified belief. This class provides an introduction to the relevant philosophicalfoundational issues underlying ML research. It involves both philosophical reflection and exercises providing concrete experience with ML methods. A unifying theme is that strong definitions of successful learning imply correspondingly strong limits on what can be learned.
Prerequisites: none listed
Courseweave needs JavaScript to show sections, workloads and planning tools.