Scalability and Precision Improvement of Neural Program Synthesis

Abstract

Mosts of the neural synthesis construct encoder-decoder models to learn a probability distribution over the space of programs. Two drawbacks in such neural program synthesis are that the synthesis scale is relatively small and the correctness of the synthesis result cannot be guaranteed. We address these problems by constructing a framework, which analyzes and solves problems from three dimensions: program space description, model architecture, and result processing. Experiments show that the scalability and precision of synthesis are improved in every dimension.