Breaking reversibility accelerates Langevin dynamics for non-convex optimization

Xuefeng Gao; Mert Gürbüzbalaban; Lingjiong Zhu

Conference paper, 2020

Breaking reversibility accelerates Langevin dynamics for non-convex optimization

Xuefeng Gao, Mert Gürbüzbalaban, Lingjiong Zhu

Advances in Neural Information Processing Systems (NeurIPS), 33, pp. 17850–17862, 2020.

Publisher page arXiv:1812.07725

In brief

Langevin dynamics explores a non-convex landscape by adding noise to gradient descent, and the usual version is reversible in time. Adding a carefully chosen non-reversible drift leaves the stationary distribution unchanged but speeds up both convergence to it and escape from local minima, and the paper quantifies the resulting gains for non-convex optimization and for the generalization of the solutions found.

Topics

Sampling & Langevin algorithms

Cite

@inproceedings{gao2020reversibility,
  title   = {Breaking reversibility accelerates Langevin dynamics for non-convex optimization},
  author  = {Xuefeng Gao and Mert Gürbüzbalaban and Lingjiong Zhu},
  year    = {2020},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume  = {33},
  pages   = {17850--17862},
  url     = {https://proceedings.neurips.cc/paper/2020/hash/cebd648f9146a6345d604ab093b02c73-Abstract.html},
  note    = {arXiv:1812.07725},
}