Figure 1.

Contrastive learning as presented in SimCLR [39]. The input images x_i and x_j are augmented via random transformations, embedded via encoder f to yield embedding h, and then projected via g to yield projection z. The contrastive objective maximizes the similarity between positive pair z_i1 and z_i2 as well as positive pair z_j1 and z_j2 (i.e., “attract”) while minimizing the similarity between all other (negative) pairs (i.e., repel).