Hedging Against Uncertainty via Multiple Diverse Predictions

What does a young child or a high-school student with no knowledge of probability do when faced with a problem whose answer they are uncertain of? They make guesses.

Modern machine perception algorithms (for object detection, pose estimation, or semantic scene understanding), despite dealing with tremendous amounts of ambiguity, do not.

In this talk, I will describe a line of work in my lab where we have been developing machine perception models that output not just a single-best solution, rather a /diverse/ set of plausible guesses. I will discuss inference in graphical models, connections to submodular maximization over a “doubly-exponential” space, and how/why this achieves state-of-art performance on Pascal VOC 2012 segmentation dataset. Following my own advice, I will talk about talk some other cool things as well (including deep learning of course).