Hacker News new | ask | show | jobs
by fsavard 2788 days ago
I confess I haven't read the whole thing but that figure 4 looks more like a set of approaches, rather than a continuum as in the periodic table.

Their figure 1, which appears to describe fundamental design tradeoffs, makes me think of CAP theorem "maps", like this:

https://www.researchgate.net/figure/CAP-Theorem_fig4_2214620...

But the whole idea of figuring out first principles and "fundamental" design choices is quite appealing. They have a section saying that a big goal here would be automatic design of optimal data structures. They talk of machine learning techniques (they mention Bayesian optimization and reinforcement learning).

That seems a very interesting research direction. In the same vein there's this talk by Jeff Dean about how they use ML at Google to replace all sorts of heuristics in data structure design (bloom filters etc.) with machine learning to optimize performance, though from what I recall it doesn't automatically change the algorithm itself:

http://learningsys.org/nips17/assets/slides/dean-nips17.pdf

(discussed on HN previously https://news.ycombinator.com/item?id=15892956)

EDIT: I think they cite a paper by Dean and others which was part of that talk

1 comments

> I confess I haven't read the whole thing

This is a sort of functional bug in forums such HN.

Commenting on paper like this requires time to digest and reflect on the content. For a conceptual paper such as OP I would need at least a day (to figuratively sleep on it) but returning a couple of days later to make (hopefully) informed comment is no longer interactive. There is archival and future reference value, of course, but then rebuttals to possible misconceptions will not accompany the comment.

It's quite funny I made a very similar lament a while ago:

https://news.ycombinator.com/item?id=16230668

I only post when I think I can somehow add value to the discussion, at least value for some people who might read the comment. Obviously some people with in-depth knowledge of the paper and field in general will probably already be aware of the link I posted (for example), but others might find it an interesting avenue to explore. I know I often come to HN comments for this reason: a topic seems interesting, I want to learn of related stuff and/or get opinions from people with deeper expertise.

So even if it's only for "future reference", it might actually bring value _for some_. At least that's what I tell myself :)

Yeah, that and tracking long-running topics are things old-fashioned forums (phpBB style) are better at than the newer, more ranking-oriented ones (HN, reddit, ...), since old discussions can be pushed up again.
Although long forum threads only really work if someone regularly updates the first entry with newest infos.