Adversarial Examples Are Not Bugs, They Are Features with Aleksandar Madry

Today we're joined by Aleksander Madry, Faculty in the MIT EECS Department, a member of CSAIL and of the Theory of Computation group. Aleksander, whose work is more on the theoretical side of machine learning research, walks us through his paper "Adversarial Examples Are Not Bugs, They Are Features," which was published previously presented at last year's NeurIPS conference. In our conversation, we explore the idea of adversarial examples in machine learning systems being features, with results that might be undesirable, but still working as designed. We talk through what we expect these systems to do, vs what they're actually doing, if we're able to characterize these patterns, and what makes them compelling, and if the insights from the paper will inform opinions on either side of the deep learning debate.