Amit Karmon: 'Machine Learning and the New Riddle of Induction'
Chair: Dan Gallagher
Faculty Respondent: Bernhard Salow
Learning machines seem to be able to learn about data they have not yet seen from running algorithms on distinct data they get as inputs. This seems highly analogous to what we do in our inductive reasoning. If the problems of induction are general problems for learning, then a suitable version of them should apply to learning machines. There are reasons to think that we can learn a lot about the problems of induction by examining whether and how they apply to learning machines. Learning machines are much simpler than humans, we know much more on how they operate, and they can (at least in principle) be mathematically modelled.
In this talk, I examine whether the New Riddle of Induction (aka “the Grue Problem”) can indeed be suitably modified to apply to learning machines. I argue that it can’t. The problem quite immediately dissolves when we try to translate it to learning machines. If time permits, I’ll discuss how this failure to translate the problem to machines sheds light on how the New riddle can be dissolved for humans as well.
