Loss function perspective of machine learning: What does a machine learn?
Machine learning has become so complicated that it often lacks a clear perspective to understand what a system indeed learns. In today's talk, I would like to briefly introduce machine learning from the viewpoint of loss functions and optimization, which has a fundamental connection to how one should evaluate a system. The central question in the talk is, "Does our system capture what we want to elicit?" Through our recent studies on robust machine learning and self-supervised learning, I show some examples of what a system does and does not learn, which will be an important step towards reliable knowledge induction.