We show that a generative random field model, which we call generative
ConvNet, can be derived from the commonly used discriminative ConvNet, by
assuming a ConvNet for multi-category classification and assuming one of the
categories is a base category generated by a reference distribut