Over the past decades, we have seen machine learning making great strides in understanding visual scenes and natural languages. Yet, most of its success relies on training models on a massive amount of data offline and evaluating them in a similar test environment. By contrast, humans can learn new concepts and skills with very few examples. In order to approach the human ability of quick learning in the natural world, we need to think beyond the classic machine learning paradigm. In this talk, I will present some new learning tasks and highlight key ingredients of meta-learning and representation learning to make machines learn more naturally with limited labeled data.