Approximating mutual information of highdimensional variables using learned representations