What is zero-shot learning used for?#
It dramatically reduces data collection costs and allows an AI system to become useful much faster. Your business can adapt to new products, customer issues, or market trends without waiting for a lengthy and expensive retraining process for your AI tools.
How does zero-shot learning work?#
Instead of only being trained on examples (like pictures of cats and dogs), the AI is also trained on rich descriptions or attributes associated with those examples (e.g., a dog "is a canine, barks, has a tail"). This teaches the AI to build a deep understanding of the features that define a category, rather than just memorising what the category looks like.
When the system later encounters an entirely new category it has never seen—for instance, a "llama"—it is provided only with a description of that new category ("is a camelid, has long woolly fur, native to South America"). The AI uses its learned knowledge of attributes to match the new item to the new description, allowing it to correctly identify the llama without ever having seen one before.
How is zero-shot learning different from few-shot learning?#
Zero-shot learning enables an AI to identify new categories without any prior examples of those specific categories. This is distinct from traditional AI training (supervised learning), which requires hundreds or thousands of labelled examples for every category. It also differs from "few-shot learning," a middle ground where the AI is given just a handful of examples of a new category to learn from.