Neural Program Learning from Input-Output Examples

Опубликовано 19 апреля 2017, 23:16

Most deep learning research focuses on learning a single task at a time - on a fixed problem, given an input, predict the corresponding output. How should we handle problems where the task is not known beforehand? One potential framework to address this issue is program learning, in which the neural network produces outputs conditioned upon inputs and a program specification. This is an appealing approach because it allows generalization to new, unseen tasks by simply providing a new program specification at inference time. In this talk, I will discuss two neural approaches to program learning when program specifications take the form of input-output examples: neural program synthesis and neural program induction. I will discuss the strengths and weaknesses of each approach. I will discuss empirical results on a string transformation program learning benchmark, FlashFill, as well as several architectural modifications we made to support the structure of the problem. I will end by discussing promising avenues for future research in neural program learning, and some challenges which remain.

See more on this video at microsoft.com/en-us/research/v...