Recently, I went through an ICLR paper SELF-LABELLING VIA SIMULTANEOUS CLUSTERING AND REPRESENTATION LEARNING. In the paper, authors discussed simultaneously labeling the images and training a network using these labels that can be termed as a chicken-and-egg problem: we require the labels to train the network, and we require the network to predict the labels.
Can we generalize it as: we require responses to learn a model, but using a model for unsupervised learning through representation learning and feature extraction we can predict the responses? So that it will be useful for learning problems based on unlabeled data set with few labels or almost no labels due to some shortcomings.
