I have mentioned this twice before on HN, and I have been making progress: I am interested in creating "compact" (self-defined, non-standard term) machine learning models. A compact model is a smaller version of a model that in getting its size reduced tries to lose out on as little accuracy as possible.

I have a library now, that implements a subset of my promising ideas so far [1]. I am not done polishing and testing the implementation yet - hence, version < 1. And I am working on some more ideas - so more theory and code to (hopefully) come soon.

[1] https://compactem.readthedocs.io/