Differentiation of Blackbox Combinatorial Solvers

ICLR 2020 (spotlight)

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

Embed blackbox combinatorial solvers into neural networks without any sacrifices! How to turn c++ code for solving e.g. travelling salesman into a differentiable NN building block. Mostly theoretical work with update rule resembling classical SVM-based methods but revamped to make good theoretical sense for deep learning. The start of #blackboxbackprop.

This paper was among the top rated papers at OpenReview for the entire ICLR 2020 conference.

Links: Arxiv OpenReview Github Blogpost ICLR (video)