Variational inference is a technique for approximating intractable distributions by optimizing a tractable variational distribution. It was used by Infomedia to identify global events from Twitter data by separating tweets into topics using latent Dirichlet allocation (LDA). Initially Gibbs sampling for LDA took nearly a day but variational inference using Gensim's LDA model converged much faster in 2 hours. Variational inference works by choosing a family of distributions and minimizing the Kullback-Leibler divergence between the true posterior and the variational distribution. This can be done using coordinate ascent variational inference or stochastic variational inference for large datasets.