What is zero-shot learning?
Zero-shot learning (ZSL) is a machine learning scenario where an AI model is trained to recognize and categorize objects or concepts without seeing examples of those specific categories during training.
Unlike traditional supervised learning that requires many labeled examples, ZSL enables models to generalize to new classes without explicit training on those classes. This approach is handy when obtaining labeled data for all possible courses, such as rare diseases or newly discovered species, is impractical or impossible.
How does zero-shot learning work?
Zero-shot learning works by leveraging auxiliary information and transfer learning to predict unseen classes. The critical aspects of how ZSL functions include:
- Auxiliary Information: ZSL uses additional semantic information about classes, such as textual descriptions, attributes, or embedded representations, to understand unseen classes.
- Transfer Learning: Models often repurpose knowledge from pre-trained models to minimize training time and resources.
- Semantic Embeddings: Both classes and samples are represented as vector embeddings in a shared semantic space, allowing for comparison between different data types.
