Zero-Shot Learning (ZSL) is a paradigm of machine learning where a model is capable of recognizing classes for which no training examples are available.
Formally, two sets of classes are considered: the training classes \( \mathcal{Y}_{seen} \) and the test classes \( \mathcal{Y}_{unseen} \) with \( \mathcal{Y}_{seen} \cap \mathcal{Y}_{unseen} = \emptyset \).
The goal is to learn a function \( f: \mathcal{X} \to \mathcal{Y}_{unseen} \) that correctly assigns input data \( x \in \mathcal{X} \) to an unknown class by accessing a semantic representation \( \phi(y) \) of the classes.
These representations can be attributes, word vectors, or other semantic embeddings that are available for both seen and unseen classes.
A common approach is to project input data and class attributes into a shared space to compute similarities and thus assign unknown classes.
Definition:
“Zero-Shot Learning is a learning method where a model recognizes new classes based on their semantic descriptions without having training data for these classes.”
Source:
Palatucci, M., Pomerleau, D., Hinton, G. E., & Mitchell, T. M. (2009). Zero-Shot Learning with Semantic Output Codes. Advances in Neural Information Processing Systems (NeurIPS).