Sparse coding is currently an active topic in signal processing and pattern recognition. Meta Face Learning (MFL) isatypical sparse coding method and exhibits promising performance for classification. Unfortunately, due to using the l1-norm minimization, MFLis expensive to compute and is not robust enough. To address these issues, this paper proposesafaster and more robust version of MFLwith the l2-norm regularization constraint on coding coefficients. The proposed method is used to learnaclass-specific dictionary for facial expression recognition. Extensive experiments on two popular facial expression databases, i.e., the JAFFEdatabase and the Cohn-Kanade database, demonstrate that our method shows promising computational efficiency and robustness on facial expression recognition tasks.