Hacker News Reader

We have been working on this recently on the Google Brain team.

We are working both on synthesizing programs from scratch (see https://arxiv.org/abs/2002.09030 for example) and on understanding computer programs using machine learning (see e.g. https://arxiv.org/abs/1911.01205).

I'm always happy to correspond with people about these topics.

View on HN

Have you considered attempting to develop a "differentiable programming language" so that it's possible to do gradient descent directly on the space of programs? There was some work on building a "differentiable lambda calculus", https://core.ac.uk/download/pdf/82396223.pdf, but I haven't heard of anyone trying to use something like that in modern program synthesis.

Late on the thread but you might be interested in this: https://www.tensorflow.org/swift

That's something different: it allows doing gradient descent on the space of paramaters for programs, not on the space of programs.

Love to chat. Sent u an email (former HFT and kernel guy).

One topic which is feel is neglected is a good GCN (or any GNN) to operate on existing code trees. Most approaches seem to prefer seq or at most tree inputs.

Is this simply not finding yet a good network architecture, or is it a performance issue ?

I dont see any email listed on your profile. Would like to have a chat to work/learn on this stuff

Reply on news.ycombinator.com