News for hackers. Stuff that matters.
Re:Open Problems in Computational geometry by ogogmad on Saturday May 17, 2025 @11:16AM (#44014936)
Is any of the new machine learning tech promising here? I recall some new invariants of minimal surfaces were discovered only a few years ago by a DeepMind-made AI - and that's before LLMs. I'm wondering if AI can invent notions as powerful as homology groups: It could go about this by constructing lossy compressors whose outputs can still be used to accurately predict properties of geometric objects. That is what homology groups and the like are for.
Re:Open Problems in Computational geometry by jebarker on Saturday May 17, 2025 @12:28PM (#44015383)
This does seem like one math domain where there's some potential for program synthesis approaches like the recent AlphaEvolve and others. I say that because some of these problems you could feasibly write automatic evaluation code and solve them by the LLM spitting out a constructor for solutions and then doing hill climbing. That's not true in many areas of math though. There's also problems here that require a proof and maybe would be approachable eventually using automated theorem proving. But there's also problems that don't obviously fit into either of those categories.
Re:Open Problems in Computational geometry by bubblyworld on Saturday May 17, 2025 @01:14PM (#44015662)
A technicality is that those compressors need to provably predict properties of said geometric objects to be interesting to mathematicans (a statement which itself needs appropriate encoding in mathematical language - what exactly you want probably depends on the problem you hope to solve). Something that superficially looks good but works as a black box wouldn't cut it.

My understanding of ML interpretability research at the moment is that it's very early days - too early, perhaps, to hope for a general purpose compressor algorithm that produces proofs of it's own workings (that seems like the holy grail). But of course I would love to be wrong!