Matching networks, in general, propose a framework that learns a network that maps a small training dataset and tests an unlabeled example in the same embeddings space. Matching networks aim to learn the proper embeddings representation of a small training dataset and use a differentiable kNN with a cosine similarity measure to check whether a test data point has already been seen.
Matching networks are designed to be two-fold:
- Modeling level: At the modeling level, they propose matching networks, which uses advances made in attention and memory that enable fast and efficient learning.
- Training procedure: At the training level, they have one condition—the distribution of training and test sets must be the same. For example, this could mean showing a few examples per class and switching the task from minibatch to minibatch, similar to how...
