Despite the recent remarkable advances in deep learning, we are still far from building machines with human-like general intelligence, for instance, understanding the world in a fast, structured, and generalizable way. The dominant stream in contemporary AI hopes to achieve human-level performance via purely data-driven methods, \ie, fitting deep neural networks on a massive amount of training data. However, these methods are often trapped in a dilemma of “big data, small tasks'', and are hard to interpret and generalize.