On Discriminative Probabilistic Modeling for Self-Supervised Representation Learning