Contrastive Learning Objectives constitute a class of optimization strategies in self-supervised representation learning wherein models maximize mutual information between paired image-text modalities while minimizing it for non-matching pairs to achieve semantic alignment. Formally defined by the InfoNCE loss function or its variants, this mechanism operates within the domain of multimodal deep learning as a specific instantiation of metric learning and contrastive regularization. The core theoretical principle dictates that embedding spaces are constructed such that semantically related image-text pairs converge in high-dimensional latent space distance while distractors diverge, establishing an invariant geometric structure independent of specific dataset distributions.
Explains contrastive representation learning via SimCLR: augmenting an image twice, pulling the pair together and pushing other samples apart with the InfoNCE/NT-Xent loss, and why the learned featur…