In this paper we present the nonconvex exterior-point optimization solver (NExOS)—a novel first-order algorithm tailored to constrained nonconvex learning problems. We consider the problem of minimizing a convex function over nonconvex constraints, where the projection onto the constraint set is single-valued around local minima. A wide range of nonconvex learning problems have this structure including (but not limited to) sparse and low-rank optimization problems. By exploiting the underlying geometry of the constraint set, NExOS finds a locally optimal point by solving a sequence of penalized problems with strictly decreasing penalty parameters. NExOS solves each penalized problem by applying a first-order algorithm, which converges linearly to a local minimum of the corresponding penalized formulation under regularity conditions. Furthermore, the local minima of the penalized problems converge to a local minimum of the original problem as the penalty parameter goes to zero. We implement NExOS in the open-source Julia package NExOS.jl, which has been extensively tested on many instances from a wide variety of learning problems. We demonstrate that our algorithm, in spite of being general purpose, outperforms specialized methods on several examples of well-known nonconvex learning problems involving sparse and low-rank optimization. For sparse regression problems, NExOS finds locally optimal solutions which dominate glmnet in terms of support recovery, yet its training loss is smaller by an order of magnitude. For low-rank optimization with real-world data, NExOS recovers solutions with 3 fold training loss reduction, but with a proportion of explained variance that is 2 times better compared to the nuclear norm heuristic.
Citation
@article{NExOS, title={Exterior-point Optimization for Nonconvex Learning}, author={Das Gupta, Shuvomoy and Stellato, Bartolomeo and Van Parys, Bart P.G.}, journal={arXiv preprint arXiv:2011.04552}, note={\url{https://arxiv.org/pdf/2011.04552.pdf}}, year={2020} }