What is the relationship between zero-shot learning and few-shot learning?

Zero-shot learning (ZSL) and few-shot learning (FSL) are both techniques designed to address scenarios where labeled training data is scarce, but they differ in how they leverage available information. ZSL aims to handle tasks where no labeled examples are provided for a target class during training. Instead, it relies on auxiliary information—like textual descriptions, semantic attributes, or relationships between classes—to generalize to unseen categories. For example, a ZSL model trained to recognize animals might infer that a “zebra” has stripes and four legs, even if it was never explicitly shown zebra images during training. In contrast, FSL uses a small number of labeled examples (e.g., 5–20 instances) to adapt to a new task. For instance, a few-shot image classifier might learn to distinguish between five new bird species after seeing just three labeled images per species.

The key distinction lies in the amount of supervision and the mechanisms for generalization. ZSL requires prior knowledge about how classes relate to one another, often through metadata or pre-trained embeddings. This makes it useful in cases where gathering labeled data is impossible, such as classifying rare medical conditions with no available patient images. FSL, on the other hand, assumes minimal labeled data is accessible and focuses on efficiently extracting patterns from those examples. A common FSL approach involves meta-learning, where a model is pre-trained on many tasks to quickly adapt to new ones with limited data. For example, a language model fine-tuned on a handful of customer support emails can learn to categorize new emails without extensive retraining. Despite their differences, ZSL and FSL can overlap: a zero-shot model might use a few examples to refine its predictions, blending both approaches.

Choosing between ZSL and FSL depends on the problem constraints. ZSL is ideal when labeled data for target classes is entirely unavailable but semantic or relational data exists. FSL is better suited when a small dataset can be curated, even if it’s imperfect. In practice, developers often combine these methods. For instance, a pre-trained vision-language model like CLIP can perform zero-shot image classification by matching images to text descriptions but can also be fine-tuned with a few examples to improve accuracy. Both approaches prioritize efficient generalization, but ZSL emphasizes external knowledge, while FSL focuses on maximizing limited labeled data. Understanding their trade-offs helps developers design systems that adapt to real-world data scarcity.