This paper proposes an adaptive face recognition algorithm to jointly classify and learn from unlabeled data. It presents an efficient design that specifically addresses the case when only a single sample per person is available for training. A dictionary composed of regional descriptors serves as the basis for the recognition system while providing a flexible framework to augment or update dictionary atoms. The algorithm is based on l1 minimization techniques and the decision to update the dictionary is made in an unsupervised mode via non-parametric Bayes. The dictionary learning is done via reverse-OMP to select atoms that are orthogonal or near orthogonal to the current dictionary elements. The proposed algorithm was tested with two face databases showing the capability to handle illumination, scale, and some moderate pose and expression variations. Classification results as high as 96% were obtained with the Georgia Tech database and 94% correct classification rates for the Multi-PIE database for the frontal-view scenarios.