On the Informativeness of Supervision Signals

Learning transferable representations by training a classifier is a well-established technique in deep learning (e.g., ImageNet pretraining), but it remains an open theoretical question why this kind of task-specific pre-training should result in ''good'' representations that actually capture the underlying structure of the data. We conduct an information-theoretic analysis of several commonly-used supervision signals from contrastive learning and classification to determine how they contribute to representation learning performance and how the dynamics of learning are affected by training parameters such as the number of labels, classes, and dimensions in the training dataset. We validate these results empirically in a series of simulations and conduct a cost-benefit analysis to establish a tradeoff curve that enables users to optimize the cost of supervising representation learning on their own datasets.