0 Members and 2 Guests are viewing this topic.
DEEPCODER: LEARNING TOWRITEPROGRAMSMatej Balog∗Department of EngineeringUniversity of CambridgeAlexander L. Gaunt, Marc Brockschmidt,Sebastian Nowozin, Daniel TarlowMicrosoft ResearchABSTRACTWe develop a first line of attack for solving programming competition-style prob-lems from input-output examples using deep learning. The approach is to train aneural network to predict properties of the program that generated the outputs fromthe inputs. We use the neural network’s predictions to augment search techniquesfrom the programming languages community, including enumerative search andan SMT-based solver. Empirically, we show that our approach leads to an orderof magnitude speedup over the strong non-augmented baselines and a RecurrentNeural Network approach, and that we are able to solve problems of difficultycomparable to the simplest problems on programming competition websites.