Efficient Learning of Restricted Boltzmann Machines Using Covariance estimates