Matching Networks for One Shot Learning

ABSTRACT

Deep learning algorithms typically require vast amounts of data to achieve high performance, contrasting sharply with human ability to learn new concepts from a single example. We introduce Matching Networks, a novel neural architecture designed to tackle the problem of one-shot learning. The model leverages an attention mechanism over a learned representation of a small support set of examples to predict the label of an unobserved query example. By employing a non-parametric approach combined with end-to-end differentiable deep neural networks, Matching Networks can rapidly adapt to new, unseen classes without any fine-tuning. We achieve state-of-the-art results on the Omniglot dataset and the challenging miniImageNet dataset.