Zero-Shot Learning is the ability of a model to perform a task or recognize a category it was never explicitly trained on. By leveraging semantic relationships and auxiliary information (like text descriptions), the model can generalize its knowledge to unseen scenarios.
A longstanding goal in AI to mimic human flexibility. The concept has roots in the 2000s, with significant formalization by Lampert et al. (2009) in computer vision.
Popularized by CLIP (OpenAI) which can classify any image given a text description, and GPT-3 which performs tasks via simple instructions without fine-tuning.