Contrastive learning

The loss for an anchor $a$, positive $p$ and negatives $n_j$ is $-\log \frac{e^{s(a,p)/\tau}}{e^{s(a,p)/\tau} + \sum_j e^{s(a,n_j)/\tau}}$ with temperature $\tau$. What counts as a positive (paraphrase, query-answer pair, augmented view) defines the geometry learned, and the quality of negatives matters as much as their number.