A Neural Framework for Low-Shot Learning

Abstract

There has been a growing interest in developing machine learning models that are capable of low- shot learning, the machine learning problem of learning from little data. Progress on low-shot learning has important practical applications to domains in which data for training state-of-the-art algorithms are scarce, for example in identifying rare diseases in medical images or personalizing online services to a user’s activity. Improvements on this task would also have important theoretical implications, as a successful solution in low-shot learning would likely also push the boundary in representation learning, natural language or image understanding, etc. Matching networks are a recently proposed model for low-shot learning that combine neural networks with nonparametric models. They were shown to perform well on benchmark low-shot learning tasks, highlighting the potential of this approach. To better understand the strengths and shortcomings of this family of models, in this work, we compare matching networks and several variants, against a strong baseline when applied to a diverse set of tasks. We find that on relatively simple low-shot learning tasks such as character recognition, specialized low-shot models are not necessary to do well. On more complex tasks such as facial recognition, we see significant improvements in accuracy when using matching networks.