---

Zero shot Learning

Zero-shot learning is a problem setup, where a model has to predict the class of samples, which belong to a novel category that was not observed during training. For example, the model could be asked to predict the presence of a zebra, having only seen horses during training.

Typically, those tasks are solved by relating attributes of the novel class to attributes of learned classes. A zebra for example, can be described to the model as a horse with black and white stripes.

This auxiliary information can be given by users or taken from databases like Wikipedia.

An example for zero-shot learning using relative similarities between the observed and unobserved classes is presented in Relative Attribute Rank#Zero-shot learning.