First, the concept of training step or epoch does not apply, since given the scale images do not need to be reused. As shown in Figure 2, the reconstruction error during training, smoothly converges to good local minima before the dataset is used up. Second, learning without using any ...