Zero-shot learning(ZSL)
Zero-shot learning is a sub-field of ML that deals with classification. In a typical classification task trains on data which is a set of mappings of inputs and their corresponding classes. Sometimes the training dataset may not consist of all classes of inputs. For example, we want to classify images of consumer products into their respective classes such as books, shirts etc. Our training set may not capture all classes of products. The problem -- how does a model trained on such a dataset classify products belonging to classes not present in the training data.
Such scenarios are quiet common in modern AI applications that deal natural language, autonomous driving, etc. In all such examples, the data comes from distributions with long tails. It could also be that there are a large number of classes(some possibly unknown) that it is practically difficult to capture all of them in the training data.
领英推荐
ZSL methods aim to solve this problem. Ever wondered how chatbots such as ChatGPT answer reasonably well to all our vague questions.
--