[HN Gopher] Statistics Postdoc Tames Decades-Old Geometry Problem
       ___________________________________________________________________
        
       Statistics Postdoc Tames Decades-Old Geometry Problem
        
       Author : rbanffy
       Score  : 159 points
       Date   : 2021-03-02 13:10 UTC (9 hours ago)
        
 (HTM) web link (www.quantamagazine.org)
 (TXT) w3m dump (www.quantamagazine.org)
        
       | tpoacher wrote:
       | "To the surprise of experts in the field, a postdoctoral
       | statistician has solved ..."
       | 
       | Wth? Am I the only one bothered by this opening statement? A
       | post-doc IS an expert in the field. One who has spent a
       | significant number of years doing a PhD to become THE expert in
       | their own particular subfield, and presumably many more years
       | after that diving into even greater detail.
       | 
       | Since when have we grown so accustomed to postdocs that we're
       | treating them as on par with undergraduates in terms of academic
       | value, and are so super surprised when they make an important
       | discovery?
        
         | Fomite wrote:
         | As others have mentioned, and as someone who does statistics, I
         | think you're putting the emphasis on the wrong word in
         | "postdoctoral statistician".
         | 
         | If I solved a major open problem in geometry, Quanta could
         | easily write "To the surprise of experts in the field, a
         | professor of epidemiology has solved..." and they'd be correct
         | in doing so.
        
         | ttymck wrote:
         | Is it not referring to the fact that they are a postdoctoral
         | _statistician_ and not a "geometrician" or something of that
         | nature?
        
           | omaranto wrote:
           | Geometer, not geometrician, but yes, you're right.
        
         | strangeloops85 wrote:
         | I think you may be misunderstanding the implication. It's not
         | that he's a postdoc that's surprising, but that he's a
         | statistician who happens to have solved a major problem in
         | convex geometry. The following quote sums it up:
         | 
         | "Chen is not a convex geometer by training -- instead, he is a
         | statistician who became interested in the KLS conjecture
         | because he wanted to get a handle on random sampling. "No one
         | knows Yuansi Chen in our community," Eldan said. "It's pretty
         | cool that you have this guy coming out of nowhere, solving one
         | of [our] most important problems.""
         | 
         | The great thing is, the reaction of the community working on
         | these type of problems was to understand and verify the result.
         | And once it was, all credit to him.
        
           | tpoacher wrote:
           | Thanks for this explanation. That makes it slightly better I
           | guess. It still sounds subtly presumptive though, when placed
           | as the subheading. It feels a bit like when journalists
           | report female scientists and feel the need to report their
           | marital status in the first paragraph.
        
             | bopbeepboop wrote:
             | It's not like that at all.
             | 
             | It's like reporting "chemist solves problem in particle
             | physics" -- which is interesting and a bit unexpected, but
             | not at all an unrelated factoid about the person in the way
             | marital status would be.
        
             | gnulinux wrote:
             | Absolutely not at all. Statistics and geometry are very
             | different fields. Sure they both use some part of
             | mathematics and an undergrad level statistics and geometry
             | curriculum will have many common classes, but at post-doc
             | level it is definitely surprising. I'm not saying a
             | statistician is not qualified to do research in geometry --
             | they might be competent enough to do so -- it's just
             | surprising because statistics and geometry work on
             | different problems.
        
         | necubi wrote:
         | The surprising thing is not that he's a post-doc, it's that
         | he's a statistician.
        
         | ska wrote:
         | Others have pointed out the misread here, but I'll add that
         | typically the thing you are actually expert in after this
         | amount of study is typically very narrow. So when they say "the
         | surprise of experts in the field" on many topics "the experts"
         | they are referring to is often a few dozen people.
        
         | a1369209993 wrote:
         | > A post-doc IS an expert in the field.
         | 
         | Other people have already mentioned it, but to phrase it
         | another way: he's a expert in a _different_ field.
        
       | OmicronCeti wrote:
       | Basically everything quanta writes is so good. They have some of
       | the best science writers around in my opinion, and their efforts
       | to make science and discovery accessible to non-experts and the
       | public is admirable. I highly recommend their podcast that breaks
       | down some of the popular stories, as well as "The Joy of X" which
       | are long-form interviews with leading scientists.
        
         | strangeloops85 wrote:
         | Completely agree! I've really enjoyed Erica Klarreich's writing
         | in particular. I think she may very well be one of the most
         | gifted and capable writers today on math in particular.
        
           | angry-tempest wrote:
           | She interviewed just the right people too! She is like the
           | equivalent of Ed Yong for math
        
             | strangeloops85 wrote:
             | It helps (and perhaps is no surprise, given her easy
             | facility with the material she covers) that she's a Math
             | PhD as well :)
        
         | melling wrote:
         | Funded by billionaire mathematician and hedge fund founder,
         | where the algorithms make all the decisions
         | 
         | https://m.youtube.com/watch?v=QNznD9hMEh0
         | 
         | 1976 winner of the Oswald Veblen Prize in Geometry
         | 
         | https://en.m.wikipedia.org/wiki/Oswald_Veblen_Prize_in_Geome...
        
           | thechao wrote:
           | The person referenced above is "James Simon".
        
             | superbcarrot wrote:
             | Simons*
             | 
             | Mostly known for being the founder of Renaissance
             | Technologies which is one of the largest and most
             | successful hedge funds.
        
               | angry-tempest wrote:
               | Also the Simons Foundation, which funds the Simons
               | Institute, which is just fantastic
        
         | thomasahle wrote:
         | They're entertaining to read, but I've never come away from one
         | with any sort of understanding of the actual math and how I
         | might use it. It would help if they at least defined the
         | problem and the new theorem.
        
           | mcguire wrote:
           | Did they define " _substantial area_ " and I just missed it?
        
             | btilly wrote:
             | You missed it.
             | 
             |  _Bourgain guessed that some of these lower-dimensional
             | slices must have substantial area. In particular, he
             | conjectured that there is some universal constant,
             | independent of the dimension, such that every shape
             | contains at least one slice with area greater than this
             | constant._
             | 
             | In other words there exists C > 0 such that if the
             | n-dimensional hypervolume of an n-dimensional convex shape
             | is 1, then there must be an n-1'th dimensional slice of
             | n-1-dimensional hypervolume at least C.
             | 
             | What was proven is weaker. For any e > 0 there is an N such
             | that if N < n, then any n-dimensional convex shape of
             | n-dimensional hypervolume 1 must have an n-1'th dimensional
             | slice of n-1-dimensional hypervolume at least 1/n^e.
             | 
             | It turns out that the exact things that were proven are
             | good enough to improve our bounds on how quickly various
             | machine learning algorithms will converge. Which means we
             | aren't just hoping based on how they worked in a few
             | examples, we have a theory explaining it.
        
             | mbeex wrote:
             | No, and in particular by speaking about slices with a lower
             | dimension d-1 it becomes not immediately clear, what their
             | volume has to do with the one in dimension d of the
             | original body, mentioned at the begin of the article.
        
           | btilly wrote:
           | They actually did a better job of that with this problem than
           | I usually see.
           | 
           | But if you don't have a substantial background, it may be
           | hard to track.
        
           | MaxBarraclough wrote:
           | I figure this is because they're trying to make it
           | accessible. As a somewhat mathematically literate non-
           | mathematician I doubt I'd get much out of the real formulae
           | that stump today's researchers. Quanta do a good job at
           | dumbing it down enough to make sense to someone like me
           | without making it completely empty. More specifically, Erica
           | Klarreich does a good job of it - she seems to write most if
           | not all of the Quanta articles that end up on the HN front
           | page.
        
       | revel wrote:
       | Aside from being a wonderful achievement and advancement, this is
       | a really excellent article. It's very impressive to read such
       | technical material presented in such a beautiful way.
        
         | leafmeal wrote:
         | Quanta is free, and virtually all of their articles read like
         | this.
        
       | zellyn wrote:
       | The final paragraph was beautiful...
        
       | paulpauper wrote:
       | It seems like unsolved problems in math are being solved at an
       | increasingly fast rate. I think a combination of the internet
       | making information more readily accessible combined with having
       | more people alive to work on such problems, are contributing
       | factors to this.
        
       | tovej wrote:
       | As far as I can tell, this is a relevant proof when searching
       | through high-dimensional parameter spaces (e.g. machine
       | learning).
       | 
       | This would mean that overall geometry is not a (big) factor when
       | it comes to getting stuck on local optima. Depending on how a
       | random walk is implemented however, the conditions might create a
       | practically concave search space.
        
         | 6gvONxR4sf7o wrote:
         | When it comes to getting stuck at a local optimum, I think it's
         | the convexity of the loss function that matters, not just the
         | convexity of the parameter space. As I understand it, this
         | result says that for convex losses, some simpler samplers work
         | near enough to ideally.
        
         | scythmic_waves wrote:
         | > This would mean that overall geometry is not a (big) factor
         | when it comes to getting stuck on local optima.
         | 
         | I thought the results only applied to convex shapes. Search
         | spaces in ML need not be convex, right? Or am I missing
         | something?
        
           | tovej wrote:
           | I assume they are convex, typically cuboid. This is the type
           | of space you get when each parameter is searched for a
           | certain range.
           | 
           | Surely there could be search spaces that aren't convex. In
           | that case the range of a variable would depend on the values
           | of other variables. If you have an example of such a case I'd
           | be interested in knowing about it.
        
             | Bootvis wrote:
             | If you have linear constraints the search space will be
             | convex.
        
             | scythmic_waves wrote:
             | Ah I was confusing the loss landscape with the parameter
             | search space because you said "local optima". Yep, I
             | imagine most parameter search spaces are convex.
        
       ___________________________________________________________________
       (page generated 2021-03-02 23:01 UTC)