We present a method for learning sparse representations shared across multiple tasks. This method is a generalization of the well known single task 1-norm regularization. It is based on a novel non-convex regularizer which controls the number of learned features common across the tasks. We prove...