07:06 < jenkins-mlpack> Starting build #1935 for job mlpack - svn checkin test (previous build: SUCCESS)

07:39 < jenkins-mlpack> Project mlpack - svn checkin test build #1935: SUCCESS in 33 min: http://big.cc.gt.atl.ga.us:8080/job/mlpack%20-%20svn%20checkin%20test/1935/

07:39 < jenkins-mlpack> andrewmw94: Rectangle Tree Traversal implementation.

13:30 < oldbeardo> naywhayare: I just sent you the latest version of the code, please review it

15:37 < naywhayare> oldbeardo: ok, I will get to it probably tomorrow

15:39 < oldbeardo> okay

15:45 < oldbeardo> naywhayare: I have a question

15:45 < oldbeardo> if you look at point 2 of Lemma 1

15:47 < oldbeardo> the ExtractSVD() function returns the decomposition of A * V_hat * V_hat'

15:48 < oldbeardo> how do we get the decomposition of the original matrix?

16:06 < naywhayare> oldbeardo: what do you mean, for testing?

16:08 < oldbeardo> no, the cosine tree does not construct A_hat = A * V_hat * V_hat'

16:09 < oldbeardo> I mean the obtained basis is not successful in reconstructing the original matrix

16:11 < oldbeardo> the Monte Carlo estimation algorithm especially doesn't make sense to me

16:23 < naywhayare> I am going to have to wait to look into this

16:23 < naywhayare> but I will see what I can find out later

16:41 < Anand_> Marcus : I have added the tests. Have a look. It seems like we are done with class conversions for all metrics!

16:45 < marcus_zoq> Anand: Okay, great ... do you mind if I look at it tomorrow?

16:50 < Anand_> No. It will be fine!

17:40 < andrewmw94> naywhayare: what is the secondBound in neighbor_search/neighbor_search_stat.hpp

17:50 < naywhayare> andrewmw94: basically, that's an undocumented nightmare. I'm sorry you've come across it

17:50 < andrewmw94> ahh.

17:50 < naywhayare> let me see what that bound actually is

17:50 < andrewmw94> do I want to know more?

17:51 < naywhayare> well, probably? if you take a look at the tree-independent dual-tree algorithms paper, there's this incredibly complex B(N_q) function defined in the section on nearest neighbor search

17:51 < naywhayare> so the firstBound and secondBound statistics are cached information that allows fast calculation of B(N_q)

17:51 < naywhayare> this is all used in NeighborSearchRules::CalculateBound()

17:52 < naywhayare> however, it is possible that the comments at the top of the function are out of date. I have not yet gotten to my NeighborSearchRules overhaul because it's still very experimental code

17:52 < andrewmw94> right

17:52 < naywhayare> according to my comments, and I think this is actually correct, FirstBound is a bound on "the worst candidate distance of any descendants of this node."

17:53 < andrewmw94> I got that part. I couldn't understand the second comment though

17:53 < naywhayare> SecondBound is a bound on "the worst distance of any descendants of this node assembled using the best descendant candidate distance modified using the furthest descendant distance."

17:53 < naywhayare> yeah

17:53 < naywhayare> I just wrote that and I am not sure I understand it completely either

17:53 < naywhayare> the paper assembles B(N_q) from two functions B_1(N_q) and B_2(N_q)

17:54 < naywhayare> the second bound is based on the triangle inequality

17:54 < naywhayare> so suppose there's some descendant point with nearest neighbor candidate distance x

17:55 < naywhayare> and call that descendant point p_i. then, thanks to the triangle inequality, it must be true that every other descendant point p_j has a nearest neighbor with distance less than or equal to (x + d(p_i, p_j))

17:55 < naywhayare> and d(p_i, p_j) is bounded by 2 * furthestDescendantDistance

17:55 < naywhayare> maybe that makes more sense?

17:55 < andrewmw94> ahh, now it does

17:56 < naywhayare> ok. I am going to have to find some way to write better comments, because that comment on SecondBound is definitely WTF

17:56 < andrewmw94> "assembled" is a bit weird there, but I'm not sure what I would call it