arxivst stuff from arxiv that you should probably bookmark

Catalyst Acceleration for Gradient-Based Non-Convex Optimization

Abstract · Mar 31, 2017 17:27 ·

stat-ml math-oc

Arxiv Abstract

  • Courtney Paquette
  • Hongzhou Lin
  • Dmitriy Drusvyatskiy
  • Julien Mairal
  • Zaid Harchaoui

We introduce a generic acceleration scheme to accelerate gradient-based convex optimization algorithms to solve possibly nonconvex optimization problems. The proposed approach extends the Catalyst acceleration for convex problems and allows one to venture into possibly nonconvex optimization problems without sacrificing the rate of convergence to stationary points. We present promising experimental results for sparse matrix factorization and for learning neural networks.

Read the paper (pdf) »