Abstract
We present a new method for evaluating and training unnormalized densitymodels. Our approach only requires access to the gradient of the unnormalizedmodel's log-density. We estimate the Stein discrepancy between the data densityp(x) and the model density q(x) defined by a vector function of the data. Weparameterize this function with a neural network and fit its parameters tomaximize the discrepancy. This yields a novel goodness-of-fit test whichoutperforms existing methods on high dimensional data. Furthermore, optimizing$q(x)$ to minimize this discrepancy produces a novel method for trainingunnormalized models which scales more gracefully than existing methods. Theability to both learn and compare models is a unique feature of the proposedmethod.