Advanced Deep Learning with Python

Zero-shot learning

We'll begin with zero-shot learning (k = 0), where we know that a particular class exists, but we don't have any labeled samples of that class (that is, there is no support set). At first, this sounds impossible—how can we classify something we have never seen before? But in meta learning, this is not exactly the case. Recall that we leverage knowledge of previously learned tasks (let's denote them with a) over the task at hand (b). In that regard, zero-shot learning is a form of transfer learning. To understand how this works, let's imagine that a person has never seen an elephant (another highly unlikely example), yet they have to recognize one when they see a picture of it (new task b). However, the person has read in ...