Human-level concept learning through probabilistic program induction

• Motivation:After seeing only one example, human can generalize the example as a concept, distinguish the example from other examples which belong to different classes, and generate new examples from this class while deep learning models require lots of training data to finish the above tasks.
• Method: The paper uses the handwritten characters dataset and chooses primitives from the library as subparts and then they are used to form parts. Parts and relations make simple programs. Running the programs, we can new tokens.
• Contribution: This paper is a science paper, very rare in this field. In fact, I still do not understand some points in this paper but this paper did integrate into models the human learning process.