Contrastive Learning is the self-supervised representation learning paradigm that trains encoders to pull together representations of semantically similar inputs (positive pairs) and push apart representations of dissimilar inputs (negative pairs) — learning powerful visual and multimodal features from unlabeled data that transfer effectively to downstream tasks through linear probing or fine-tuning.

Core Mechanism:

• Positive Pair Construction: two augmented views of the same image form a positive pair; augmentations (random crop, color jitter, Gaussian blur, horizontal flip) create views that differ in low-level appearance but share high-level semantics — forcing the encoder to capture semantic similarity rather than pixel-level features
• Negative Pairs: representations of different images serve as negatives; the contrastive objective pushes positive pairs closer than any negative pair in the embedding space; quality and diversity of negatives significantly impact learning quality
• InfoNCE Loss: L = -log(exp(sim(z_i, z_j)/τ) / Σ_k exp(sim(z_i, z_k)/τ)) where z_i, z_j are positive pair embeddings and z_k includes all negatives; temperature τ (0.05-0.5) controls the sharpness of the distribution over similarities
• Projection Head: encoder output is mapped through a small MLP (2-3 layers) to the contrastive embedding space; only the encoder output (before projection) is used for downstream tasks — the projection head absorbs augmentation-specific information

Method Evolution:

• SimCLR (2020): simple framework using large batch sizes (4096-8192) for negative pairs; batch normalization across GPUs provides implicit negative mining; demonstrated that augmentation design and projection head nonlinearity are critical design choices
• MoCo (2020): momentum-contrast maintains a queue of negatives from recent batches, decoupling negative set size from batch size; momentum encoder (slowly updated copy of the main encoder) provides consistent negative representations; enables contrastive learning with standard batch sizes (256)
• BYOL (2020): eliminates negatives entirely using a predictor network and stop-gradient — online network predicts the target network's representation; momentum target prevents collapse; proved that contrastive learning doesn't strictly require negatives
• DINO/DINOv2 (2021/2023): self-distillation with no labels using multi-crop strategy and Vision Transformer backbone; student network matches teacher network's centered and sharpened output distribution; discovers emergent semantic segmentation without any segmentation supervision

Design Choices:

• Augmentation Strategy: the most critical hyperparameter; augmentation must be strong enough to force semantic-level learning but not so strong that it destroys class-discriminative information; color distortion + random crop + Gaussian blur is the standard recipe
• Batch Size vs Queue Size: SimCLR requires large batches (4096+) for sufficient negatives; MoCo decouples with a queue (65536 negatives); BYOL/DINO avoid the issue entirely by eliminating negatives
• Encoder Architecture: ResNet-50 was the standard backbone; ViT-based encoders (DINOv2) achieve significantly better representations with emergent properties (spatial awareness, part discovery); encoder choice affects both representation quality and transfer performance
• Training Duration: contrastive pre-training typically requires 200-1000 epochs (vs 90 for supervised ImageNet); longer training consistently improves representation quality with diminishing returns beyond 800 epochs

Evaluation and Transfer:

• Linear Probing: freeze the encoder, train only a linear classifier on labeled data; measures representation quality independent of fine-tuning capacity; DINOv2 ViT-g achieves 86.5% ImageNet accuracy with linear probing — close to full fine-tuning results
• Few-Shot Learning: contrastive representations enable strong few-shot classification (>70% accuracy with 5 examples per class on ImageNet); the learned similarity metric generalizes across domains and tasks
• Dense Prediction: contrastive pre-training produces features useful for detection and segmentation; DINOv2 features exhibit emergent correspondence and segmentation properties without any pixel-level supervision

Contrastive learning is the breakthrough that made self-supervised visual representation learning practical — enabling models trained on unlabeled image collections to match or exceed supervised pre-training quality, reducing the dependence on expensive labeled datasets and establishing the foundation for vision foundation models.