HanbaekLyu / SRMM

Stochastic Regularized Majorization-Minimization

SRMM

Stochastic Regularized Majorization-Minimization (SRMM)

Hanbaek Lyu,
"Stochastic regularized block majorization-minimization with weakly convex and multi-convex surrogates" (arXiv 2023)

Stochastic majorization-minimization (SMM) is a class of stochastic optimization algorithms that proceed by sampling new data points and minimizing a recursive average of surrogate functions of an objective function. We propose an extension of SMM called Stochastic Reguarlized Majorizaiton-Minimizaiton (SRMM) where surrogates are allowed to be only weakly convex or block multi-convex, and the averaged surrogates are approximately minimized with proximal regularization or block-minimized within diminishing radii, respectively.

In this repository, we provide a special version of SRMM proposed in the reference below, where prox-linear surrogates with proximal regularization is used. The resulting algorithm is equivalent to the following iterates

Setting $\lambda=0$ reduces it to the following `double-averaging PSGD', which was first investigated by Nesterov and Shikhman in 2015:

CIFAR-10 Demo using DenseNet-121 and ResNet-34:

File description

1. src.SRMM.py : main algorithm source file that implements the SRMM algorithm in (23)
2. demos.cifar10: contains scripts for generating the DenseNet and ResNet training/testing accuracies figure

Author

• Hanbaek Lyu - Initial work - Website

License

This project is licensed under the MIT License - see the LICENSE.md file for details