Miltos

I've been working on a research machine learning-based tool (link: http://groups.inf.ed.ac.uk/naturalize/ ) tool that analyzes source code identifiers and makes suggestions for renaming them. The goal is to reduce unnecessary...

The current [2D Map](https://ml4code.github.io/tsne-viz.html) is okay, but could be improved visually as well as with better filtering. * For example, filter by tag. * Coloring of datapoints. * Better tooltips...

This is not necessarily wrong, but I want to point out that using a ReLU [here](https://github.com/emalgorithm/structured-neural-summarization-replication/blob/b5b1883e2fbf32c65ac9d5813c6b260811c294cd/models/lstm_decoder.py#L18) is not a very common choice as far as I know. This might not...

(suggested by @bzz)

The current model may return `%UNK%` that's not useful. Filter them. (suggested by @bzz)

Some of the suggestions of the machine learning model may be incorrect. Use a (configurable?) type checker to remove incorrect suggestions.

Get information from the statistics of reactions (e.g. for autotuning confidence thresholds, improving training etc.)

Create and use the full Typilus model instead of graph2class. - [ ] Implement it in `ptgnn` - [ ] Use action cache to store intermediate result - [ ]...

When the model predicts with very high probability that an annotation is different from the original one, add a comment.