Great question—this is such a common point of confusion because the two fields overlap heavily in their end goal (learning embedding spaces where similar samples cluster close together and dissimilar ones stay far apart). Let’s break down their core distinctions clearly:

1. Primary Goal & Use Case Focus #

• Metric Learning: It’s inherently task-driven. The core objective is to design or learn a distance metric that directly enables a specific downstream task (e.g., image retrieval, face verification, few-shot classification). The embedding space is optimized explicitly for how well it performs that single task’s distance-based requirements.
• Contrastive Learning: It’s centered on generalizable representation learning. The goal is to learn a universal embedding space that captures rich, transferable features from data—often without relying on manual labels. This pre-trained embedding can then be fine-tuned for a wide range of downstream tasks, not just one.

2. Supervision Signal Source #

• Metric Learning: Relies on explicit, human-provided supervision. This can take the form of:
  • Pairwise labels (e.g., "these two samples are similar/dissimilar")
  • Class labels (used to generate positive/negative pairs or triplets)
  • Domain-specific constraints (e.g., in person re-identification, same person across different cameras = positive pair)
• Contrastive Learning: Uses self-generated or weak supervision. Most state-of-the-art methods leverage data augmentation to create positive pairs (different views of the same sample) and treat all other samples in a batch/queue as negative pairs. No manual labeling is required for pre-training, making it ideal for unlabeled datasets.

3. Training Paradigm #

• Metric Learning: Typically operates on small batches with carefully constructed sample groups (e.g., triplets of anchor-positive-negative). The loss function focuses on optimizing the relative distances between these pre-defined groups to enforce task-specific constraints.
• Contrastive Learning: Thrives on large batches (or even external negative sample queues, like in MoCo). The loss encourages the model to distinguish between the augmented views of a single sample and all other unrelated samples in the batch, emphasizing global feature distinctiveness rather than local triplet/pair constraints.

4. Key Example Algorithms #

• Metric Learning: Triplet Loss, Siamese Networks (with supervised contrastive loss), Mahalanobis Distance Learning
• Contrastive Learning: SimCLR, MoCo, BYOL, SimSiam

Quick Summary #

To put it simply: Metric Learning is about customizing a distance metric for a specific task, while Contrastive Learning is about learning a versatile embedding space that works across many tasks. They share the "similar close, dissimilar far" principle, but their starting points, supervision methods, and end use cases set them apart.

内容的提问来源于stack exchange，提问作者JustinGong