Inspired by humans' remarkable ability to master arithmetic and generalize to unseen problems, we present a new dataset, HINT, to study machines' capability of learning generalizable concepts at three different levels: perception, syntax, and semantics. In particular, concepts in HINT, including both digits and operators, are required to learn in a weakly-supervised fashion: Only the final results of handwriting expressions are provided as supervision. Learning agents need to reckon how concepts are perceived from raw signals such as images (ie, perception), how multiple concepts are structurally combined to form a valid expression (ie, syntax), and how concepts are realized to afford various reasoning tasks (ie, semantics). With a focus on systematic generalization, we carefully design a five-fold test set to evaluate both the interpolation and the extrapolation of learned concepts. To tackle this challenging problem, we propose a neural-symbolic system by integrating neural networks with grammar parsing and program synthesis, learned by a novel deduction–abduction strategy. In experiments, the proposed neural-symbolic system demonstrates strong generalization capability and significantly outperforms end-to-end neural methods like RNN and Transformer. The results also indicate the significance of recursive priors for extrapolation on syntax and semantics.