positive and negative pairs

Every contrastive method rests on a single design decision: which samples count as positives (should be similar) and which count as negatives (should be different). In self-supervised instance discrimination the rule is mechanical, a positive pair is two augmented views of the same source image, and negatives are views of all other images. In supervised contrastive learning the definition is richer, all images sharing a class label are positives of each other. The choice of this rule silently determines what the representation will treat as the same and what it will separate.

Negatives carry most of the difficulty and most of the leverage. Easy negatives (a cat versus a totally unrelated sky) are quickly satisfied and contribute almost nothing to the gradient once separated. Hard negatives (a cat versus a similar-looking dog, or a different crop of a visually similar scene) sit close in embedding space and produce the large, informative gradients that actually sharpen the boundaries. Methods therefore care about negative mining, deliberately surfacing or up-weighting hard negatives, and about simply having many negatives, since more candidates raise the chance that some are hard.

The dangerous failure mode is the false negative: a sample labelled negative that is actually semantically identical to the query, for example two different photographs of the same dog breed treated as negatives under instance discrimination. The loss then actively pushes apart things that ought to be together, injecting noise. This false-negative problem is one of the strongest motivations for the non-contrastive family (BYOL, SimSiam, SwAV, DINO), which avoids explicit negatives entirely, and for clustering-based methods that group likely-similar images before contrasting.