The document discusses human-level concept learning through probabilistic program induction. It compares the principles of learning from big data versus human learning. Humans can learn rich concepts from just one or a few examples by using compositional and causal reasoning, as well as learning-to-learn abilities. The document introduces Bayesian Program Learning (BPL), which uses these principles to perform one-shot learning on the Omniglot data set. BPL is evaluated using visual Turing tests and one-shot classification performance to demonstrate human-level concept learning from limited examples.