For large data it can be very time consuming to run gradient based optimization, for example to minimize the log-likelihood for maximum entropy models. Distributed methods are therefore appealing and
