A NIPS paper got accepted!

I’m happy to say that our paper “SEGA: Variance Reduction via Gradient Sketching” about variance reduction for coordinate descent methods was accepted to NIPS. In this work, we discover how coordinate descent methods can be extended to problems with non-separable regularizer. We consider minibatch, importance sampling and acceleration of the produced method. What’s more, we additionally introduced a variant with non-trivial sketches of the gradient, under maximal possible stepsizes can be larger by an arbitrary number than in classical coordinate descent methods.

Konstantin Mishchenko
Konstantin Mishchenko
Postdoctoral Researcher

I study optimization and its applications in machine learning.