Now that we have understood how to use a relation network in one-shot and few-shot learning tasks, we will see how to use relation networks in a zero-shot learning setting where we will not have any data points under each class. However in zero-shot learning, we will have meta information which is information about the attributes of each class and that will be encoded into the semantic vector, , where the subscript c represents the class.
Instead of using a single embedding function for learning the embeddings of support and query sets, we use two different embedding functions, and respectively. First, we will learn the embeddings of the semantic vector, , using and learn the embeddings of the query set, , using . Now we will concatenate these embeddings using our concatenation operation, :
Then, we will feed this result to the relation function and compute the relation score as follows: