verne.freenode.net changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
< naywhayare>
stephentu: yeah, I was planning on submitting an application
< naywhayare>
want to be a mentor? :)
< naywhayare>
we'll probably have fewer students this year... last year I mentored three or four, it was a terrible idea; took way more time than I had expected
< naywhayare>
this year I'll mentor one...
< stephentu>
naywhayare: i'm happy to be a mentor
< stephentu>
i have a few neat ideas
< stephentu>
parallel optimization
< stephentu>
is the theme
< stephentu>
(they arent novel ideas, just it'd be great to get this stuff implemented solidly)
< stephentu>
also the manifold stuff since i might not be able to get to it
< stephentu>
shit is starting to pick up
< stephentu>
also another idea is like trying to make mlpack more composable like sklearn
< stephentu>
the pipelines abstraction
< stephentu>
but thats more of a software engineering problem
< stephentu>
also mlpack is very optimization / dual tree heavy
< stephentu>
we're kind of light on the whole bayesian thing
< stephentu>
graphical model
< stephentu>
i'm not the biggest fan of that stuff, but i would be happy to mentor some implementations of say np bayes
< stephentu>
even just LDA might be useful
< naywhayare>
sure, I could agree with that
< naywhayare>
we'll have to make sure the implementations are high-quality and fast
< naywhayare>
thinking of which, I'm hoping to spend some time soon writing up more detailed style/coding guidelines, and I wouldn't mind having you (and marcus and whomever else) take a look at them too
< naywhayare>
imitating scikit's functionality is fine, but we have to make sure there is some compelling reason to use mlpack instead of scikit in those cases
< stephentu>
sure
< stephentu>
i might be more of a fan of immutability than you are
< stephentu>
but happy to take a look
< naywhayare>
the idea being that once you create an optimizer/classifier/regressor/some other machine learning object, its parameters cannot be changed?
< stephentu>
something like that. like it kind of bothers me that for the SDP object, you can just change the C and Ai matrices like that
< stephentu>
you can even make the Ais and bs inconsistent w/ each other
< stephentu>
(eg different lengths)
< naywhayare>
well, I mean, we have to assume some intelligence on the part of the user
< naywhayare>
but I do see what you mean
< stephentu>
haha
< stephentu>
but honestly after using it
< stephentu>
i have found it much easier
< stephentu>
to deal with
< stephentu>
mutating in place
< stephentu>
versus
< stephentu>
constructing all args
< stephentu>
and then using std::move
< stephentu>
its ok now, but if we want to move to parallelism
< stephentu>
having immutable objects makes reasoning about code so much easier
< stephentu>
do gsoc mentors get anything?
< stephentu>
or is it all charity
< naywhayare>
Google pays $500 per mentor plus t-shirt and possibly a trip to the mentor summit in October in San Jose (each organization gets to send two members)
< stephentu>
san jose
< stephentu>
that sounds so exciting
< stephentu>
</sarcasm>
< naywhayare>
well, for you it is not as exciting as for someone who has never been to California :)
< naywhayare>
really though, the summit is a lot of fun
< naywhayare>
even though San Jose isn't very exciting
< naywhayare>
(it is walkable though -- I had to run halfway across town to buy a laptop because mine abruptly died there last year)
< naywhayare>
(it took... about an hour to get to the nearest best buy... sigh)
< stephentu>
uber?
< naywhayare>
I was being cheap :)
< stephentu>
i'm sure its fun i was just kidding
< naywhayare>
:)
< naywhayare>
they closed down Great Adventure for us last year, it was a good time
< stephentu>
oh teh amusement park next to teh hilton?
< naywhayare>
yeah
< naywhayare>
they didn't have all the rides open, but still, it was a great time, no lines or anything
< stephentu>
ya i mean google has quite a bit of clout in the silicon valley
< naywhayare>
we didn't find anyone from scikit or any other machine learning libraries though
< stephentu>
hey which one are you
< stephentu>
i realize i have no idea what you look like
< naywhayare>
leftmost, the least european of the bunch :)
< stephentu>
nice
< naywhayare>
marcus is next to me, then the three shogun guys (one of them must have won the lottery for extra slots I guess)
< naywhayare>
the whole program is a lot of fun, but it can be a bit of a time sink depending on the student
< stephentu>
we get to filter students though right?
< stephentu>
i actually did gsoc in undergrad
< stephentu>
as a participant
< naywhayare>
the first year I mentored Marcus and one other student... Marcus basically didn't need any help from me, and the other student simply disappeared
< naywhayare>
ah, neat! for which organization?
< stephentu>
scala
< naywhayare>
and yeah, we get to select
< stephentu>
i was trying to beef up the remote actors implementation
< stephentu>
but it toko about the entire summer to get it working
< stephentu>
and then i never got to really integrate it
< stephentu>
b/c actually writing real software is much harder
< naywhayare>
last year I think we had 30-40 applications and accepted 5... of those, probably only 6-8 were actually competitive
< naywhayare>
yeah
< naywhayare>
a summer isn't very much time to really get something baked into the code
< stephentu>
ya it was way too ambitious
< naywhayare>
the CF code has gone through two summers now and still needs some more cleanup
< stephentu>
so ya id liek to see the other side of it
< naywhayare>
one of my next free-time projects was going to be to sit down with the CF code, integrate it all, and write up a nice tutorial on how to use CF with numerous different factorizations
< naywhayare>
yeah, the time contribution is a bit picky
< stephentu>
and we still need to hook up matrix completion
< stephentu>
to CF
< naywhayare>
sorry... that didn't make any sense... let me try again:
< naywhayare>
the time contribution is a bit tricky
< naywhayare>
like I said my two students in the first year needed basically nothing
< naywhayare>
so the next year I mentored four, thinking it would be about the same
< stephentu>
didtns cale
< stephentu>
*scale
< stephentu>
?
< naywhayare>
well, the students needed more help for their projects
< naywhayare>
so I was spending like 15-20 hours a week on GSoC
< stephentu>
gross
< naywhayare>
a lot got done for mlpack, but I didn't get much research done...
< stephentu>
so i was thinking since i'll prob be doign research
< stephentu>
and studying for prelims
< stephentu>
this woudl be a good way to breka up teh routine
< stephentu>
a
< stephentu>
bit
< naywhayare>
yeah
< naywhayare>
with just one (or even two students) it's pretty nice, maybe a handful of hours here or there to help them out
< stephentu>
ya i'm thinking one good student
< naywhayare>
I felt stretched really thin last summer; I didn't manage to get super-involved with any of the students
< stephentu>
who is ambitious
< naywhayare>
there's some code that I still haven't managed to look at and it's what, five months later?
< stephentu>
even better if i can recruit them from berkeley undergrads
< naywhayare>
:)
< naywhayare>
I got to meet one of the students last year... he's an undergraduate at CUA in DC, and I was passing through, so I got some lunch with him
< naywhayare>
there really aren't very many applicants from the US, generally, though
< naywhayare>
one of the reasons is that it's not too hard to pull down ~$6k or so in a summer doing just a regular internship with a software company or whatever
< stephentu>
oh ya you take a pay cut
< stephentu>
but i mean
< stephentu>
who wants to write liek
< stephentu>
another web app
< naywhayare>
it's true
< stephentu>
we're doing some cool shit here
< naywhayare>
machine learning is a bit of a niche too, so I think we see fewer of the really excited open-source people as a result
< stephentu>
actually i was at facebook though and it was a lot of fun
< stephentu>
but i got really lucky
< naywhayare>
well... maybe? I don't know what the typical applicant pool for GSoC looks like
< stephentu>
and got to hack on the compiler
< naywhayare>
that does sound like a good time
< stephentu>
but most peopel were doign boring stuff
< stephentu>
like writing php apps
< naywhayare>
in 2010 I interned at Google, with the Similar Pages team, who were having me run some big mapreduces
< naywhayare>
they took three days to run
< naywhayare>
but my team didn't have anything else for me to do
< naywhayare>
so... I played a ___lot___ of pinball
< stephentu>
lol
< stephentu>
oh internships
< stephentu>
machine learning is niche?
< stephentu>
i feel like
< stephentu>
everybody out here
< stephentu>
is doing machine learning
< naywhayare>
I mean, I asked repeatedly, "I want to hunt bugs! Give me something to do!", and I got a lot of "well, we'll look into that... have you checked out this cafe yet?"
< naywhayare>
you live in a bubble :)
< stephentu>
well i used to live in cambridge, MA
< stephentu>
and it was like that too
< naywhayare>
cambridge is also a bubble...
< stephentu>
when i was at MIT
< stephentu>
everyobyd did machine learning
< naywhayare>
yeah
< naywhayare>
MIT :)
< stephentu>
i guess i only live in bubbles
< stephentu>
i jsut assume like everybody has gone through the derivation of SVM
< stephentu>
in their free time
< naywhayare>
maybe a better way to put what I meant is that despite the fact that everyone is talking about machine learning, the set of people who can actually write and understand machine learning algorithms is quite small
< stephentu>
then why are there so many fucking papers in NIPS
< stephentu>
haha
< stephentu>
its insane
< stephentu>
i guess the few who understand
< stephentu>
really understand
< stephentu>
and crank stuff out like machines
< naywhayare>
from a NIPS perspective, people look like they are machines because they are part of a machine (a big research lab)
< naywhayare>
speaking from experience, without a lab, it's very hard to get anything published at NIPS
< naywhayare>
in part because as a one-man show it's very hard to keep your finger on the pulse of what all reviewers will say
< naywhayare>
and also because in a lab, risk is distributed because papers have many co-authors
< naywhayare>
so you can have some people working on the crazier ideas that might not work, and some others work on the solid ideas that might only be incremental
< naywhayare>
but my opinion is based on my experience, which is only my experience and a small subset of all experience, so, YMMV :)
< stephentu>
i see
< stephentu>
thats interesting
< naywhayare>
maybe? I dunno, it's certainly not guaranteed to be a well-informed opinion :)
< naywhayare>
anyway... I'm headed to bed. I hope to find some time to work on mlpack tomorrow, but I'm pretty underwater with the ICML deadline coming up
< stephentu>
later dude
stephentu_ has joined #mlpack
udit_s has joined #mlpack
stephentu_ has quit [Ping timeout: 244 seconds]
udit_s has quit [Ping timeout: 245 seconds]
udit_s has joined #mlpack
udit_s has left #mlpack []
udit_s has joined #mlpack
HeikoS has joined #mlpack
udit_s has quit [Quit: leaving]
stephentu has quit [Ping timeout: 264 seconds]
HeikoS has quit [Quit: Leaving.]
naywhayare has quit [Ping timeout: 265 seconds]
jbc_ has joined #mlpack
HeikoS has joined #mlpack
aditya1995 has joined #mlpack
HeikoS1 has joined #mlpack
HeikoS has quit [Quit: Leaving.]
aditya1995 has quit [Quit: Page closed]
aditya1995 has joined #mlpack
udit_s has joined #mlpack
naywhayare has joined #mlpack
< naywhayare>
sigh, on Friday I lose the motherboard in mlpack.org, and on Monday there's a cooling failure
aditya1995 has quit [Ping timeout: 246 seconds]
< udit_s>
naywhayare: :o so what is happening now ?
< naywhayare>
just waiting on the facilities people to fix the cooling system, then it will come back up
< naywhayare>
shouldn't be more than a few hours, I think (but I'm not sure). I'll wander over there after lunch and see if I can get an ETA
< naywhayare>
at the end of the summer I'll be moving that system to a place that should have less downtime... (and also replacing the hardware entirely so it stops failing for a few years)
< udit_s>
oh.
< udit_s>
though, why not completely move it on to the cloud ? Or am I forgetting something ?
< naywhayare>
I'm not completely against moving mlpack.org onto some other hosting service, but I don't really want to pay for anything
< naywhayare>
if there was some free alternative that would host mlpack.org and allow me to run it in the same way (i.e. SSH in, check out website with git, or something similar), I'd be happy to do that
< naywhayare>
but I just run it on that particular box because it's convenient, already running, and I use it for other stuff too (personal website, etc.)
< udit_s>
naywhayare: have you heard of the student pack of github ?