Optimizing Rank-based Metrics with Blackbox Differentiation

CVPR 2020 (oral) nomination for best paper award

Authors: Michal Rolínek*, Vít Musil*, Anselm Paulus, Marin Vlastelica, Claudio Michaelis, Georg Martius

#blackboxbackprop allows to differentiate through the ranking function out of the box. This, along with a few other tricks, allows for efficient optimization of rank-based metrics. With a minimal implementation overhead, we obtain competitive results on metric learning benchmarks and on object detection.

Links: Arxiv Github Blogpost YouTube(1min) YouTube(5min)

The first real application of #blackboxbackprop is out there! It can out-of-the-box differentiate the ranking function (just by calls to torch.argsort) and help directly optimize rank-based metrics. Simple, blazing-fast, and about on-par with SOTA! https://t.co/8O5urU7gSy pic.twitter.com/b4VWKMUtsP

— Michal Rolínek (@MichalRolinek) December 17, 2019