Metadata-Version: 2.1
Name: mice
Version: 0.1.0
Summary: Multi-iteration Stochastic Estimator
Home-page: https://bitbucket.org/agcarlon/mice
Author: Andre Gustavo Carlon
Author-email: agcarlon@gmail.com
License: GPL v3
Platform: UNKNOWN
Description-Content-Type: text/markdown
License-File: LICENSE

Multi-iteration Stochastic Estimator
============

The multi-iteration stochastic estimator (MICE) is an estimator of gradients to be used in stochastic optimization. It uses control-variates to build a hierarchy of iterations, adptively sampling to keep the statistical variance below tolerance in an optimal fashion, cost-wise. The tolerance on the statistical error decreases proportionally to the square of the gradient norm, thus, SGD-MICE converges linearly in strongly convex L-smooth functions.


