lichenglu
commented
6 years ago I will do the first one, extending the machine learning aided auto-coding feature.
Description
When users use the "select-to-code" feature, instead of showing them suggestions by usage frequency or last time used, we can train a model to give suggestion based that specific users' behaviors.
What is good about this way of using machine learning is that the model could be corrected by the user. Because if we offer the user a list of five codes, and none of them get chosen, then we know have a source of penalty for the model, which sounds like a good case for reinforcement learning.
As time goes by, ideally, each user will have a model that really understands them.
Difficulties
- Small training sets. It is very possible that users only use the app, say, once per months...and each time it
- Sounds like we need a backend for training models. But I don't think this should be a backend that we host, because it could be expensive. We need to figure out a painless way that users host their own backends for model training.
@liusashmily This is the place we can think out loud for experimental features once we have the core functionality done. Lots of popular open source projects have a separate IRC repo for this kind of discussion. But since we are weird, let's do it here. 🥇