Low-shot visual learning - the ability to recognize novel object categories from very few examples - is a hallmark of human visual intelligence. Existing machine learning approaches fail to generalize in the same way. To make progress on this foundational problem, we present a novel pr