Machine learning models today attain impressive accuracy on many benchmark tasks. But to what extent do these models generalize to the real world tasks we ultimately care about?
In this talk, we will examine the performance of current ML models from this perspective. We will start with what is perhaps one of the most striking failure modes of these models---their vulnerability to imperceptible input perturbations known as adversarial examples. We take a closer look at this brittleness, and examine how it can, in part, be attributed to the fact that our models often make decisions very differently to humans. Then, we take a step back and revisit the building blocks of the ML pipeline to identify other manifestations of such human-ML misalignments and discuss how we can make progress towards mitigating them.
This talk will be hybrid, held in-person and online on Zoom