Abstract

Deep learning for visual tasks such as image and video recognition are currently optimized in a fully-parametric manner. Prior knowledge is ignored, in favor of learning everything from scratch. In this talk, I will discuss two recent papers that aim to incorporate knowledge about the visual task at hand prior to the network optimization. First, I will discuss our recent Hyperspherical Prototype Networks, which unifies classification and regression on hyperspherical output spaces for deep learning, while utilizing prior knowledge. Second, I will discuss current work on action search in videos with prior hierarchical knowledge through prototypes in hyperbolic space. Both approaches propose to restructure the output space of deep networks in an effort to make optimization easier with the help of prior knowledge.

Bio

Pascal Mettes is an assistant professor at the University of Amsterdam. He obtained his MSc (2013) at Utrecht University, his PhD (2017) at the University of Amsterdam, was a visiting scientist at Columbia University in 2016. From 2017-2019, he was also a postdoctoral researcher at the University of Amsterdam. His research interests are in computer vision, with a focus on video understanding, learning from limited supervision, and deep learning with prior knowledge.