ABSTRACT

Traditional convolutional networks exhibit unprecedented robustness to intraclass nuisances when trained on big data. Generalization with respect to geometric transformations has been achieved via expensive data augmentation. It has been shown recently that data augmentation can be avoided if networks are structured such that feature representations are transformed the same way as the input, a desirable property called equivariance. In this talk, we show how equivariance can be realized via group convolutions, how to deal with vector and tensor fields, and how we achieve equivariance in transformers. We present results on 3D shape classification and scene reconstruction based on learning only objects but not scenes.

BIO

Kostas Daniilidis has been faculty at the University of Pennsylvania since 1998.  He is an IEEE Fellow. He was the director of the GRASP laboratory from 2008 to 2013. He obtained his Ph.D. in Computer Science from the University of Karlsruhe (now KIT) in 1992. He is a co-recipient of the Best Conference Paper Award at ICRA 2017. Kostas’ main interest today is in geometric deep learning, event-based neuromorphic vision, and their applications in vision-based manipulation and navigation.