Unrolled inference network
An iterative inference procedure in an energy-based or probabilistic model, unrolled into a fixed number of steps and trained end to end as a feed-forward network; GANs avoid this by generating in one forward pass.
An unrolled inference network (or unrolled approximate inference network) turns an iterative inference procedure into a fixed-depth neural network. In energy-based or probabilistic models such as Boltzmann machines, inference may require many mean-field or sampling steps. Those steps can be unrolled into a chain of layers and trained end to end with backpropagation, as in Goodfellow, Mirza, Courville, and Bengio’s multi-prediction deep Boltzmann machines (2013) or Stoyanov, Ropson, and Eisner’s empirical risk minimization for graphical models (2011).
GANs sidestep this pattern: the generator maps noise to a sample in a single forward pass, without running a long inference loop at training or generation time. See the GAN article and deep learning for the broader context.