E cient Learning of Sparse Overcomplete Representations with an Energy-Based Model