SampleRank: Training Factor Graphs with Atomic Gradients

We present SampleRank, an alternative to con- trastive divergence (CD) for estimating param- eters in complex graphical models. SampleR- ank harnesses a user-provided loss function to distribute stochastic gradients across an MCMC chain. As a result, parameter updates can be computed between arbitrary MCMC states. Sam- pleRank is not only faster than CD, but also achieves better accuracy in practice (up to 23% error reduction on noun-phrase coreference).