The subject of this week’s group will be the paper –

 “A unified framework for high-dimensional analysis of M-estimators with decomposable regularizers”

The formulation of an estimation problem as a regularized optimization problem is useful and familiar in ML. This formulation combines two components, a loss function that measures how well the model fits the data and some regularization function (e.g. Lasso, group Lasso, ridge regression, PCA, matrix completion etc).

The authors of this paper present a general approach for constructing consistency and convergence rates for such regularized estimators in the high-dimensional setting.

This means, when the number of parameters is in the same scale as the number of samples!

The reading may require a bit of effort, but the diligent reader will be rewarded 🙂