Efficient Second Order Online Learning by Sketching

Luo, Haipeng, Agarwal, Alekh, Cesa-Bianchi, Nicolò, Langford, John

Neural Information Processing Systems 

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. Papers published at the Neural Information Processing Systems Conference.