Low-shot Visual Recognition by Shrinking and Hallucinating Features

摘要: Low-shot visual learning---the ability to recognize novel object categories from very few examples---is a hallmark of human intelligence. Existing machine learning approaches fail generalize in the same way. To make progress on this foundational problem, we present low-shot benchmark complex images that mimics challenges faced by recognition systems wild. We then propose a) representation regularization techniques, and b) techniques hallucinate additional training examples for data-starved classes. Together, our methods improve effectiveness convolutional networks learning, improving one-shot accuracy classes 2.3x challenging ImageNet dataset.