Closed iaroslav-ai closed 8 years ago
Maybe a bit too abstract; If so let me know.
Try limiting the scope of the problem. As presently stated, it is too broad and too hard. It will be much more useful to say, "here is a link to a text - program dataset, please try to get the greatest accuracy on it", and the dataset should be pretty simple, or else the problem will be totally unsolvable. In short, try to narrow the scope to make it barely solvable.
just noticed that this pull request is similar to the paper2code one I just submitted: https://github.com/openai/requests-for-research/pull/5/files
A recent DeepMind paper http://arxiv.org/abs/1603.06744 proposed a generative model for code generation.
Code generation paper is linked to in the description of paper2code research request: https://github.com/openai/requests-for-research/pull/5/files
@yanjon thanks for the reference! @ethancaballero indeed it is similar, yours is much more detailed though. @ilyasu123 agreed. I will close this pull, and open a new one as soon as I will have something more solvable and concrete and different from what @ethancaballero proposes.
Somewhat far fetched, but maybe in some very simple setup already doable (e.g. learning to program hello world).