1. Variational Inference

    Sun 28 April 2013

    Variational Inference and Monte Carlo Sampling are currently the two chief ways of doing approximate Bayesian inference. In the Bayesian setting, we typically have some observed variables \(x\) and unobserved variables \(z\), and our goal is to calculate \(P(z|x)\). In all but the simplest cases, calculating \(P(z …

  2. Topic Models aren't hard

    Mon 21 January 2013

    In 2002, Latent Dirichlet Allocation (LDA) was published at NIPS, one of the most highly regarded conferences for research loosely labeled as "Artificial Intelligence". The next 5 or so years led to a flurry of incremental model extensions and alternative inference methods, though none have achieved the popularity of their …