The problem of variational inference is that the variational distribution is usually over-simplified and it maybe very different to the posterior distribution of interest. Stein variational gradient descent favors the stein's identity and thus using a iterative methods to make the 'variational distribution' closer to the posterior distribution of interest.