Stefan Roth

Title: Visual Learning: Knowing when we don‘t know

Abstract:
Supervised learning, particularly in the form of deep neural networks, is the workhorse of current day computer vision and has enabled incredible successes over the recent years. At the same time, deep visual learning approaches still make errors, yet one thing we have lost when moving from the previously popular graphical models to supervised deep learning is the ability to gauge when our model is not sure in its prediction. In this talk, I will review a number of recent approaches that aim to bring notions of uncertainty into standard supervised deep learning.