Optimising Distributions with Natural Gradient Surrogates