Write the basic contrastive loss and when it's used.

medium

Answer

$L\; = \;y\; \cdot \;d^{2}\; + \;(1 - y)\; \cdot \;\operatorname{max}(\mathrm{margin}\; - \;d, \;0)^{2}$
$Pulls positives together, pushes negatives apart at least by a margin.$
$Used in Siamese nets, verification (faces, signatures), and metric learning.$
$Modern SSL uses info-NCE (below) instead.$

Check yourself — multiple choice

Only cross-entropy
Pull positives together, push negatives apart by a margin — Siamese / metric learning basis
Same as MSE
Only for images

Contrastive loss: attract positives, repel negatives past a margin.

#losses#contrastive#self-supervised

214 interview questions in this topic.