Ranking & Metric Learning Losses

For tasks where the output is not a class or a number but a learned embedding space, you need losses that operate on distances or similarities between samples.

Rapid Recall

Triplet loss pulls an anchor toward a positive and pushes it from a negative with a margin, the basis of face recognition, but triplet mining is hard. Contrastive (pairwise) loss does the same for pairs. InfoNCE/NT-Xent is cross-entropy over "which item in the batch is the true match," the engine of self-supervised learning (SimCLR, MoCo, CLIP), and it loves large batches and a small temperature. Pairwise ranking losses (RankNet, LambdaRank) optimize the probability that item \(i\) ranks above item \(j\), used in search and recommendation.

§1 Triplet Loss

Symbol	Meaning
\(a\)	Anchor sample.
\(p\)	Positive, a sample of the same class/identity as the anchor.
\(n\)	Negative, a sample of a different class.
\(d(\cdot,\cdot)\)	Distance function. Usually L2 or cosine.
\(m\)	Margin, the minimum desired separation between positive and negative.

\[ L = \max\big(0, \ d(a, p) - d(a, n) + m\big) \]

Intuition: pull the anchor close to the positive, push the anchor far from the negative, with a buffer (margin). If the negative is already far enough away (\(d(a, n) \ge d(a, p) + m\)), the loss is zero, the gradient becomes inactive.

Strengths: the foundation of face recognition (FaceNet), metric learning, and embedding-based retrieval; learns useful representations without needing explicit class labels. Limitations: triplet mining is hard, most random triplets are too easy (loss = 0), so you need hard or semi-hard negative mining; slow convergence; sensitive to the margin hyperparameter.

§2 Contrastive Loss (Pairwise)

\[ \text{with } y = 1 \text{ if the pair is similar, } 0 \text{ if different:} \qquad L = (1 - y)\cdot\frac{1}{2}d^2 + y\cdot\frac{1}{2}\max(0, m - d)^2 \]

Similar pairs: pull together. Dissimilar pairs: push apart only until the margin is satisfied.

§3 InfoNCE / NT-Xent (Contrastive Learning)

Symbol	Meaning
\(z_i, z_j\)	Embeddings of a positive pair (for example, two augmentations of the same image).
\(\mathrm{sim}(\cdot,\cdot)\)	Similarity function. Usually cosine.
\(\tau\)	Temperature. Typically 0.05-0.1. Lower = sharper distribution.
\(N\)	Batch size (number of negatives drawn from the batch).

\[ L = -\log\frac{\exp(\mathrm{sim}(z_i, z_j) / \tau)}{\sum_{k=1}^{N} \exp(\mathrm{sim}(z_i, z_k) / \tau)} \]

Intuition: cross-entropy where the "classes" are which item in the batch is the true match. Pulls the embeddings of two augmentations of the same example together; pushes against everyone else in the batch.

Strengths: the foundation of self-supervised learning (SimCLR, MoCo, CLIP); no need for class labels, supervision comes from data augmentations or paired modalities (image-text in CLIP); scales well with batch size (more negatives means better signal). Limitations: requires very large batch sizes (typically 4096+) for good negatives, OR a memory bank (MoCo); sensitive to temperature \(\tau\) (typically 0.05-0.1); performance degrades if augmentations are too weak (collapse) or too strong (lose semantics).

Use when: self-supervised pretraining, multimodal learning (CLIP-style), retrieval systems.

§4 Pairwise Ranking Losses (LambdaRank, RankNet)

For learning-to-rank: compute pairwise probabilities that item \(i\) should rank above item \(j\), optimize cross-entropy on these pairs. LambdaRank weights pairs by their impact on NDCG (Normalized Discounted Cumulative Gain). Used in search and recommendation ranking, the cross-encoder reranking family used in Microsoft Bing, Seller Copilot reranking, and similar systems.