Zero-Shot / K-Shot Learning

Zero-shot learning allows a model to recognize classes it has never seen during training, while few-shot (k-shot) learning enables recognition from just a handful of labeled examples per class (typically 1-5). Both address the fundamental challenge that collecting large labeled datasets for every possible category is impractical, especially in domains where new classes appear frequently or labeling is expensive.

Zero-shot approaches typically use a shared embedding space between visual and semantic representations. CLIP maps images and text descriptions into the same space, so you can classify images by comparing their embeddings to text descriptions of candidate classes ("a photo of a cat," "a photo of a dog") without any class-specific training. Grounding DINO extends this to detection, locating objects described by free-form text. Few-shot methods learn to compare new examples against a small support set: prototypical networks compute class centroids from the support examples and classify queries by nearest centroid, while metric learning approaches learn a similarity function directly.

These capabilities are practical for rapidly prototyping new detection or classification tasks, handling long-tail distributions where some classes have very few examples, and building flexible systems that can adapt to new categories without retraining. In production, zero-shot models often serve as a starting point that gets refined with task-specific fine-tuning once enough labeled data accumulates.