This paper is great. It is like a good slice of what is going on in formal methods and machine learning right now.
- Methods of space convergence from AI, can, and do, outperform the brightest algorithm writers, especialy when the time complexity is high. And innacuracy is easy to control for.
- It is all just programming. It is research level, but still. ML algs can be used easily in code logic, if there is a layer to handle uncertainty.
- Many of the referenced papers are about software that generates code without programmers. Obviously computer vision is going to beat any human, but the best tech for other domains are referenced here. Much of it is about indirect code reasoning.
- Real data is not needed for ML (supervised) research to progress. This data, CLEVR, is synthetic. Sometimes richer synthetic datasets make a better playground.
