Zero-Shot Learning explained for advanced users

Zero-Shot Learning (ZSL) refers to methods where a model can recognize objects or classes that were not explicitly seen during training.

This is made possible by the model accessing semantic descriptions or attributes of the classes, for example text descriptions or feature vectors.

Typically, a common feature space is created in which both the known classes and the new, unknown classes can be represented.

The model learns to embed input data into this space and can thus recognize new classes based on their semantic representations.

Zero-Shot Learning is often used in image and speech recognition to make models more flexible and adaptable.