Toward Neural-Network-Guided Program Synthesis and Verification

We propose a novel framework of program and invariant synthesis called neural network-guided synthesis. first show that, by suitably designing training networks, we can extract logical formulas over integers from the weights biases trained networks. Based on idea, have implemented tool to synthesize positive/negative examples implication constraints, obtained promising experimental results. also discuss an application our method for improving qualifier discovery in ICE-learning-based CHC solving, which turn be applied verification inductive Another potential is neural-network-guided variation Solar-Lezama’s sketching.