They are jumbling the labels for the training data before giving it to the network. Thus they are misinforming the network about what it is supposed to learn.