Age estimation from faces is a challenging problem that has recently gained increasing relevance due to its potentially multi-faceted applications. Many current methods for age estimation rely on extracting computationally-demanding features from face images, and then use nonlinear regression to estimate the subject's age. This often requires matching the submitted face image against a set of face prototypes, potentially including all training face images, as in the case of kernel-based methods. In this work, we propose a super-sparse regression technique that can reach comparable performance with respect to other nonlinear regression techniques, while drastically reducing the number of reference prototypes required for age estimation. Given a similarity measure between faces, our technique learns a sparse set of virtual face prototypes, whose number is fixed a priori, along with a set of optimal weight coefficients to perform linear regression in the space induced by the similarity measure. We show that our technique does not only drastically reduce the number of reference prototypes without compromising estimation accuracy, but it can also provide more interpretable decisions.