DRASO: Declaratively Regularized Alternating Structural Optimization

Recent work has shown that Alternating Structural Optimization (ASO) can improve supervised learners by learning feature representations from unlabeled data. However, there is no natural way to include prior knowledge about features into this frame- work. In this paper, we present Declar- atively Regularized Alternating Structural Optimization (DRASO), a principled way for injecting prior knowledge into the ASO framework. We also provide some analysis of the representations learned by our method.