The reparameterization gradient has become a widely used method to obtainMonte Carlo gradients to optimize the variational objective. However, thistechnique does not easily apply to commonly used distributions such as beta orgamma without further approximations, and most practical applications of thereparameterization gradient fit Gaussian distributions. In this paper, weintroduce the generalized reparameterization gradient, a method that extendsthe reparameterization gradient to a wider class of variational distributions.Generalized reparameterizations use invertible transformations of the latentvariables which lead to transformed distributions that weakly depend on thevariational parameters. This results in new Monte Carlo gradients that combinereparameterization gradients and score function gradients. We demonstrate ourapproach on variational inference for two complex probabilistic models. Thegeneralized reparameterization is effective: even a single sample from thevariational distribution is enough to obtain a low-variance gradient.
translated by 谷歌翻译