Learning Sparse Representations by Non-Negative Matrix Factorization and Sequential Cone Programming