[PROPOSAL] Formatting. With style.

m09 commented 5 years ago

Title: Formatting. With style.
Author(s): Hugo and interested members of the ML team
Short description: Describe the current style analyzer and the future plans for a deep learning version of this same analyzer.
Categories: deep learning, unsupervised learning, meta learning, machine learning on code, machine learning, source code formatting, lookout
Deadlines: no deadline

Problem statement: why linters are not enough. What we accomplish in format analyzer (model the existing style in a repo and use it to lint new code).
A first solution: unsupervised learning. This section will describe our current analyzer. It will already be really beefy since we have many tweaks and hacks to explain. Roughly:
- Circumventing a lack of data with carefully selected features and unsupervised learning
- How to do interpretable unsupervised learning: from decision trees to rules.
- Polishing: hyper-parameter tuning, output checks with babelfish, performance analysis.
The future: meta-learning. This section will describe the plans we have to go beyond unsupervised learning and use deep learning techniques to obtain a very robust and more expressive model. Roughly:
- What is meta-learning and why is it useful here?
- Our model
- Connection to other successful approaches (auto-encoders, encoder decoder, VAEs, etc).
Conclusion, links

Management

This section will be filled by @campoy.

NOTE Please write in short lines so the review is easier to do.

campoy commented 5 years ago

This sounds amazing! Looking forward to reading the first draft 😄

vcoisne commented 5 years ago

@m09 Did you get a chance to start a draft ?

campoy commented 5 years ago

Hey, @m09 we would love to publish this ASAP to use the momentum from the release.

When do you think you'd have time to work on this?

vcoisne commented 5 years ago