Few-Shot Learning

Few-shot learning studies how models can recognize new classes or adapt to new tasks when labeled examples are scarce, a common constraint in medicine, robotics, and other specialized settings. Approaches include episodic meta-learning, metric and prototype methods, transfer from pretrained models, prompt-based and in-context learning, and augmentation or transductive refinement using unlabeled data. Recent work increasingly adapts foundation models with lightweight prompts, adapters, or no parameter updates, while also extending few-shot methods to continual learning and cross-domain transfer. A central tension is whether complex adaptation genuinely helps beyond strong pretrained representations, especially when results depend on validation choices, class imbalance, or shifts between training and test domains.