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/
jbc_ has quit [Quit: jbc_]
< naywhayare>
stephentu: what's his lack of belief stem from? the lack of scalability?
< naywhayare>
technically you could see choosing a kernel correctly as manifold learning, if it manages to map the non-linearly-separable points into a kind-of-linearly-separable set of points in the kernel space
< stephentu>
naywhayare: i think he thinks that the manifold assumption is not correct--
< stephentu>
but i sort of agree with your point of view
< stephentu>
like if you were to learn a manifold
< stephentu>
and then run a say gaussian kernel
< naywhayare>
I wouldn't be surprised if the manifold assumption isn't always right
< stephentu>
it seems quite redundant
< naywhayare>
but it only needs to be approximately right :)
< stephentu>
what i dont understand is
< stephentu>
why did it die down
< stephentu>
like does it not work well in practice
< stephentu>
did neural nets take over everything
< naywhayare>
manifold learning? it died because you can't apply it to more than like 10k points
< stephentu>
ah
< stephentu>
so scalability
< stephentu>
kind of like sdps
< stephentu>
haha
< stephentu>
why do i always like things that dont scale
< naywhayare>
well, I think there is a solution out there somewhere :)
< naywhayare>
but I mean, scalability has (kind of) been achieved for kernel methods
< stephentu>
how so?
< naywhayare>
embeddings, for instance
< naywhayare>
Nystroem approximations (sample your input points)
< stephentu>
do these techniques not work for manifold learning
< naywhayare>
sampling often doesn't work for manifold learning because of the lack of out-of-sample extensions
< naywhayare>
so if I put in new points, it's sometimes difficult (depending on the method) to map it to the unfolded manifold or whatever
< naywhayare>
there is a paper that suggests how this can be done, but I don't remember the exact details
< stephentu>
i see
< stephentu>
the fact that you arent really learning an f
< stephentu>
but just mapping hte point sin 1:1
< stephentu>
interesting
< stephentu>
question: how were you going ot use MVU in yoru research
< naywhayare>
at the time, I wasn't sure what I wanted my research to be
< stephentu>
ah, so i'm like you 3 years ago now
< stephentu>
lol
< stephentu>
any advice
< naywhayare>
5 years :(
< naywhayare>
:)
< naywhayare>
my advice is just to find something that you enjoy
< naywhayare>
performing a random walk through the field until you find something sufficiently interesting is a good way to do it, in my opinion
< naywhayare>
you get good breadth like that
< stephentu>
haha thats pretty much been what i'm doing
< naywhayare>
anyway, one of my interests was to find out how to scale nonlinear dimensionality reduction techniques
< naywhayare>
but after failing to get LRSDP+MVU working I kind of moved different ways, because I simultaneously noticed some improvements and nice abstractions for problems like nearest neighbor search
< naywhayare>
it kind of snowballed from there, and then suddenly I had more ideas than time...
< stephentu>
cool
< stephentu>
in his papers I think he says he uses some variant of LRSDP
< stephentu>
i guess he had some magic sauce?
curiousguy13 has joined #mlpack
curiousguy13 has quit [Ping timeout: 244 seconds]
< naywhayare>
I have his old code
< naywhayare>
it's incomprehensible
< naywhayare>
it's actually in the git history if you dig far enough, under like u/nvasil/mvu/ or something like that
< naywhayare>
anyway, I think that maybe he got it to converge for one particular set of parameters once
< naywhayare>
I don't think the results are bullshit but I think, as you've noticed, that LRSDP is so finicky and not well understood that it's basically unreproducible without the magical set of parameters and penalty schedule
< naywhayare>
I never got his original code to converge either...
< naywhayare>
at the same time, it wasn't documented at all, so I may have been just using it wrong
curiousguy13 has joined #mlpack
curiousguy13 has quit [Ping timeout: 272 seconds]
naywhayare has quit [Ping timeout: 240 seconds]
naywhaya1e has joined #mlpack
stephent1 has joined #mlpack
stephentu has quit [Ping timeout: 244 seconds]
jbc_ has joined #mlpack
dhfromkorea has joined #mlpack
dhfromkorea has quit [Remote host closed the connection]
karthikabinav has joined #mlpack
jbc_ has quit [Quit: jbc_]
jbc_ has joined #mlpack
jbc_ has quit [Ping timeout: 265 seconds]
vedhu63w has joined #mlpack
karthikabinav has quit []
Kallor has quit [Remote host closed the connection]
< vedhu63w>
Hi! Does mlpack contain test data on which i could probably check the results of the algorithms
< stephent1>
vedhu63w: not like sklearn has
< stephent1>
vedhu63w: i assume you're looking for something like MNIST
jbc_ has joined #mlpack
< vedhu63w>
or maybe the one that developer uses for testing?
< stephent1>
for testing, each module typically has a suite of test cases
< stephent1>
and those typically have toy datasets
< stephent1>
you can see in the tests folder if you are curious
< vedhu63w>
ohh! thanks
jbc_ has quit [Ping timeout: 245 seconds]
vedhu63w has quit [Quit: Leaving]
vedhu63w has joined #mlpack
Kallor has joined #mlpack
jbc_ has joined #mlpack
stephent1 has quit [Ping timeout: 264 seconds]
jbc_ has quit [Ping timeout: 245 seconds]
babel42 has joined #mlpack
< babel42>
hello
Kallor has quit [Remote host closed the connection]
jbc_ has joined #mlpack
jbc_ has quit [Ping timeout: 265 seconds]
jbc_ has joined #mlpack
< zoq>
babel42: Hello!
jbc_ has quit [Ping timeout: 272 seconds]
jbc_ has joined #mlpack
< babel42>
Hi zoq. I'm interested in participating in gSoC'15 can you help me get started?
jbc_ has quit [Ping timeout: 252 seconds]
< zoq>
babel42: The best way to get started is to download mlpack and compile it from source, then use it for some simple machine learning tasks. The tutorials might prove helpful: http://www.mlpack.org/tutorial.html
< zoq>
Once you've got a basic feel for mlpack programs and source, you can take a look at the list of open tickets you might find something interesting: https://github.com/mlpack/mlpack/issues.
< zoq>
Most of them are marked with a difficulty, so that might help you figure out some issues that you can handle. We are always interested in new algorithms so if you interested in some special field I think we can figure something out.
< zoq>
Also, be aware that Google hasn't selected orgs yet. We've participated in the past years, but this is no guarantee they'll select us again.
< babel42>
I like ML and am doing a course o it this semester, so It thought this would be good practice
< zoq>
Yeah, It's a great opportunity.
Kallor has joined #mlpack
babel42 has quit [Ping timeout: 246 seconds]
jbc_ has joined #mlpack
jbc_ has quit [Ping timeout: 256 seconds]
Kallor has quit [Remote host closed the connection]
jbc_ has joined #mlpack
jbc_ has quit [Ping timeout: 265 seconds]
hritikj has joined #mlpack
hritikj has quit [Ping timeout: 246 seconds]
jbc_ has joined #mlpack
jbc_ has quit [Ping timeout: 272 seconds]
jbc_ has joined #mlpack
stephentu has joined #mlpack
jbc_ has quit [Quit: jbc_]
< stephentu>
naywhaya1e: ever thought of using travis CI for builds?
< stephentu>
then we can hook up each PR
< stephentu>
to see if the tests pass
< naywhaya1e>
stephentu: thought of it, yes, had time to set it up, no :(
< naywhaya1e>
definitely a good idea to test each PR
< stephentu>
naywhayare: i have had some experience w/ travis CI
< stephentu>
i can take a stab at it
< naywhayare>
sure, feel free :)
< naywhayare>
do you need any permissions or anything?
< stephentu>
naywhayare: we'll see
< stephentu>
i'll ping you
< stephentu>
naywhayare: but alas, today i must do a pset