How do few-shot learning and zero-shot learning differ?

Few-shot learning and zero-shot learning are methods to train machine learning models when labeled data is scarce, but they differ in how they leverage available information. Few-shot learning involves training a model with a small number of labeled examples (e.g., 1–10 samples per class) for tasks it needs to perform. The goal is to generalize from these limited examples by leveraging prior knowledge from related tasks or a larger pre-trained model. In contrast, zero-shot learning requires no labeled examples for the target task. Instead, it relies on auxiliary information—like textual descriptions, semantic attributes, or relationships between classes—to infer how to handle unseen data. The key distinction lies in the presence (few-shot) or absence (zero-shot) of task-specific labeled data during training.

A few-shot learning example is image classification where a model pre-trained on a broad dataset (e.g., animals) is fine-tuned with a handful of images of rare species (e.g., a “red-crowned crane”) to recognize them. Techniques like metric learning (e.g., Siamese networks) or prompt-based tuning in language models help the model adapt quickly. For zero-shot learning, consider a text classifier trained to recognize sentiment in English reviews. Without any German examples, it might infer German sentiment by mapping German text to a shared embedding space learned during pre-training, using cross-lingual semantic relationships. Another example is object detection where a model identifies an unseen class (e.g., “electric scooter”) using textual descriptions like “two-wheeled vehicle with a battery” instead of labeled images.

Choosing between the two depends on data availability and problem constraints. Few-shot learning works when minimal labeled data can be collected, making it practical for niche domains like medical imaging with rare conditions. However, performance hinges on the quality and representativeness of the few examples. Zero-shot learning suits scenarios where obtaining labeled data is impossible, such as classifying emerging product categories in e-commerce using textual metadata. Its success relies on robust auxiliary data and the model’s ability to generalize abstract relationships. Developers should consider trade-offs: few-shot methods may require fine-tuning infrastructure, while zero-shot approaches demand careful design of semantic representations to bridge seen and unseen classes.