Efficient Second Order Online Learning - VowpalWabbit/vowpal_wabbit GitHub Wiki

OjaNewton is a sketched variant of a second order online learning algorithm called Online Newton Step (ONS). It overcomes the quadratic running time of ONS by performing the Oja's updates to keep a small sketch of the covariance matrix of gradients in a sparse manner.

Example of using OjaNewton in VW:

vw  --OjaNewton --sketch_size=10 --alpha_inverse=1.0  -d train_file  -f model_file

Here sketch_size is the number of directions that we keep for the covariance matrix (default is 10) and alpha_inverse can be viewed as a learning rate (default is 1.0).

Then, to predict from the trained model and a new data set:

vw -i model_file -d data_file -p predict_file