Meta-Learners – learning how to learn

The goal of the learner (model) is to quickly learn new tasks from a small amount of new data; hence, it is sometimes called a fast learner. When an image in the target set belongs to a new unseen class, the matching network model treats it like any other image. This implies that the initial set of network parameters should be sensitive to changes in the new tasks; small changes in the parameters will result in large improvements on the loss function of any task.

Source: blog.fastforwardlabs.com