Metric Learning
Metric learning trains an embedding where distance reflects semantic similarity, enabling retrieval and verification.
Learning what close means
Metric learning builds a representation in which the distance between two points reflects how similar they are in a task-relevant sense: same-identity faces are near, different faces are far. Instead of predicting a fixed set of classes, it learns a geometry, which lets the model handle new classes at test time by comparison, essential for face verification, image retrieval, and one-shot recognition.
Contrastive and triplet losses
The contrastive loss pulls together pairs labeled similar and pushes apart pairs labeled dissimilar beyond a margin. The triplet loss uses an anchor, a positive (same class), and a negative (different class), and requires the anchor-positive distance to be smaller than the anchor-negative distance by at least a margin. These losses optimize the embedding directly rather than a classification boundary.
Mining hard examples
Most randomly chosen triplets are already correct and give no gradient, so training stalls without hard-negative mining: selecting negatives that are close to the anchor and thus violate the margin. Semi-hard mining, choosing negatives farther than the positive but still within the margin, stabilizes training by avoiding the noisiest hardest cases.
- Enables open-set recognition: classify identities never seen in training
- Powers nearest-neighbor retrieval in the learned space
- Proxy and classification-based losses (e.g. large-margin softmax) often train more stably than triplets
- Embeddings are usually L2-normalized so cosine similarity applies
Relation to contrastive learning
Metric learning and contrastive learning share the same geometric objective; the difference is the source of pairs. Metric learning uses ground-truth labels to define similar and dissimilar, while self-supervised contrastive learning constructs pairs from augmentations. In both, the product is a reusable embedding evaluated by retrieval accuracy such as recall at k.