NeurIPS 2016poster119 citations
Efficient Second Order Online Learning by Sketching
Haipeng Luo, Alekh Agarwal, Nicolò Cesa-Bianchi, John Langford
Abstract
We propose Sketched Online Newton (SON), an online second order learning algorithm that enjoys substantially improved regret guarantees for ill-conditioned data. SON is an enhanced version of the Online Newton Step, which, via sketching techniques enjoys a running time linear in the dimension and sketch size. We further develop sparse forms of the sketching methods (such as Oja's rule), making the computation linear in the sparsity of features. Together, the algorithm eliminates all computational obstacles in previous second order online learning approaches.
BibTeX
@inproceedings{NIPS2016_15de21c6,
author = {Luo, Haipeng and Agarwal, Alekh and Cesa-Bianchi, Nicol\`{o} and Langford, John},
booktitle = {Advances in Neural Information Processing Systems},
editor = {D. Lee and M. Sugiyama and U. Luxburg and I. Guyon and R. Garnett},
pages = {},
publisher = {Curran Associates, Inc.},
title = {Efficient Second Order Online Learning by Sketching},
url = {https://proceedings.neurips.cc/paper_files/paper/2016/file/15de21c670ae7c3f6f3f1f37029303c9-Paper.pdf},
volume = {29},
year = {2016}
}