Learning what to learn in a neural program

Richard Shin, Dawn Song

Abstract: Learning programs with neural networks is a challenging task, addressed by a long line of existing work. It is difficult to learn neural networks which will generalize to problem instances that are much larger than those used during training. Furthermore, even when the learned neural program empirically works on all test inputs, we cannot verify that it will work on every possible input. Recent work has shown that it is possible to address these issues by using recursion in the Neural Programmer-Interpreter, but this technique requires a verification set which is difficult to construct without knowledge of the internals of the oracle used to generate training data. In this work, we show how to automatically build such a verification set, which can also be directly used for training. By interactively querying an oracle, we can construct this set with minimal additional knowledge about the oracle. We empirically demonstrate that our method allows automated learning and verification of a recursive NPI program with provably perfect generalization.