For this tutorial, we will be using domain-adaption meta-learning to learn a simple curve of sinusoidal data. It's a variation of model-agnostic meta-learning, but with added prior information—that is, extra relevant information about the domain is already added.
Let's begin!
Meta-learning algorithms optimize the ability of models to learn new tasks quickly. To do so, they use data collected across a wide range of tasks and are evaluated based on their ability to learn new meta-test tasks. This process can be formalized as learning a prior (that is extracting important information) over data (a range of tasks), and the fine-tuning process becomes the inference under the learned prior: ...
