Negative sampling is a training technique that uses a selected subset of negative examples rather than all possible negative classes or pairs. It improves efficiency in large classification, embedding, or representation-learning tasks while still providing contrastive learning signals.
Synonym for candidate sampling.