04:36 < naywhayare> so, I've realized that the homepage (http://www.mlpack.org/) is not very good and doesn't really provide a way for people to figure out what the library does, how to use it, where to get it, etc.

04:36 < naywhayare> I've prototyped a new homepage: http://www.mlpack.org/index-new.html , and I'd be happy to hear feedback

04:36 < naywhayare> none of the links on it work yet, this is just a basic design idea

11:18 < zoq> naywhayare: Looks really nice, perhaps I would change 'get mlpack' with 'get started'. Not sure because I like the reference to mlpack.

16:49 < naywhayare> zoq: if I change it to "get started", do you think I should move the links to "about mlpack" and "how to get help" into that section too? (it seems to me like those are things people might want to know when getting started)

16:56 < zoq> naywhayare: I guess it would make sense to move the links into the get started section.

17:01 < naywhayare> hm, but if I do that, then I don't think I have anything left for the "learn about mlpack" category

17:01 < naywhayare> I suppose I could just change it to "get started", leave the links as-is, and also have a potentially non-orthogonal "learn about mlpack" category

17:01 < naywhayare> maybe I am putting too much thought into this :)

17:03 < zoq> Maybe the sections are fine the way they are :)

17:04 < naywhayare> ok, I'll leave them as-is for now then

17:04 < naywhayare> I'll start fleshing out the rest of the new site, and then when it's all ready to go I'll let the channel know again for any other comments, then deploy it to mlpack.org

17:05 < zoq> sounds great

17:21 < naywhayare> here's an interesting post to twitter I found: https://twitter.com/michiexile/status/577739688485740545

17:21 < naywhayare> the guy is having difficulty understanding how to sample from GMMs... so someone, in the reply, instead of suggesting a statistics textbook, suggests he look at the mlpack source code :)

17:21 < naywhayare> I'm glad to know all this documentation hasn't been for nothing :)

17:24 < zoq> textbook .... just read mlpack source code :)

17:24 < naywhayare> maybe I'll find a publisher and see if they'll print some version of mlpack into a book

18:52 < KTL> not sure but GaussianDistribution::FactorCovariance uses arma::log_det ... and that logarithm appears to be base 10 ? shouldn't that be base e ? but ... it's probably just me being chaotic

19:26 < zoq> naywhayare: On OSX we need to link against libc++ instead of stdc++ at least on some 'older' versions. That should solve the homebrew build error on mountain lion as pointed out in the svd issue.

19:26 < zoq> naywhayare: https://gist.github.com/zoq/5f228a6ec890 Maybe there is a better solution?

20:09 < naywhayare> zoq: the gist link 404'ed; I don't see any other gists under your username that seem to be for the clang linker issue

20:10 < zoq> maybe now?

20:10 < naywhayare> KTL: yeah, arma::log_det uses std::log internally (see armadillo_bits/fn_log_det.hpp if you're interested)

20:11 < naywhayare> zoq: yeah, https://gist.github.com/zoq/5f228a6ec89078872108 works

20:12 < naywhayare> that looks good to me, but do you know any more about which versions of OS X require -stdlib=libc++?

20:12 < naywhayare> it looks like line 9 is missing '-lm'

20:13 < naywhayare> hang on, stupid thing for me to say... that's CXXFLAGS not linker flags

20:17 < zoq> I could reproduce the build issue on OS X Lion so OS X Lion and mountain Lion. OS X mavariks comes with an updated version of libstdc++ which covers boost fusion. Regarding http://stackoverflow.com/questions/14972425/should-i-use-libc-or-libstdc and the llvm page (http://libcxx.llvm.org/) we should generally link against libc++ on OSX.

20:21 < zoq> And of course homebrew generally builds against libc++

20:21 < naywhayare> ah, thanks for the link

20:22 < naywhayare> I don't know very much about development on OS X, so this clarifies a lot

20:24 < zoq> I guess we could say update the OS or link against libc++

20:25 < naywhayare> if you push the CMake configuration patch, then he can just use that as a patch to 1.0.12

20:25 < naywhayare> (...assuming he starts actually patching the releases right...)

20:28 < zoq> so just as create a patch or update the CMake configuration in the master branch?

20:29 < naywhayare> I'd just update it in master

20:29 < naywhayare> if he says he doesn't know how to make a patch out of that, I'll do it for him because it should be quite easy

20:35 < zoq> yeah probably the best solution

21:32 < KTL> this might be interesting to get rid of the singular covariance-matrices: http://www.doc.ic.ac.uk/~dfg/ProbabilisticInference/IDAPILecture17.pdf (the current algorithm in gaussian distribution is adding a growing perturbation to the diagonal till the determinant > 0)

21:33 < KTL> (look for "Regularisation (or shrinkage) methods")

21:36 < naywhayare> it seems like equation (9) is basically what we are doing now, and iteratively growing alpha until the matrix is nonsingular

21:36 < naywhayare> I'm still reading, but I've always thought there must be a better way

21:38 < naywhayare> if you have an idea to apply, please feel free to take a shot and submit it as a PR; I'd be more than happy to accept a better strategy

21:42 < KTL> (btw, equation 9 also lowers the entire existing covariance matrix)

21:43 < naywhayare> yeah, what we are doing isn't exactly the same, just sort of similar :)

21:43 < KTL> :D

21:44 < naywhayare> I always thought about the problem from the perspective of "can we project the matrix onto the cone of positive definite matrices in some way that minimizes || A - A' ||_F" or something like that

21:44 < naywhayare> but I didn't put too much time into it, and I couldn't quickly find any algorithms that didn't have huge amounts of overhead, which is why I eventually went with the "just add crap to the diagonal until it works" approach

21:45 < KTL> after looking through a bunch of papers i figured there is lots of academic crap around :D

21:46 < naywhayare> yeah, I figured there must be something, but I didn't have the time to devote to it

21:46 < KTL> time is limited :(