The authors aim to introduce a new method for training deep Boltzmann machines. Inspired by inference procedure they turn the model into two hidden layers autoencoder with recurrent connections. Instead of reconstructing all pixels from all (perhaps corrupted) pixels they reconstruct one subset of pixels from the other (the complement).
DBM are usually "pre-trained" in a layer-wise manner using RBMs, a conceivably suboptimal procedure. Here the authors propose to use a deterministic criterion that basically turns the DBM into a RNN. This RNN is trained with a loss that resembles that one of denoising auto-encoders (some inputs at random are missing and the task is to predict their values from the observed ones).