The generalization of representations learned via contrastive learning
depends crucially on what features of the data are extracted. However, we
observe that the contrastive loss does not always sufficiently guide which
features are extracted, a behavior that can negatively impact the