naywhayare 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/
sumedhghaisas has quit [Ping timeout: 252 seconds]
< naywhayare>
sumedh_: I don't think I have the version of the code that calculates the validation RMSE
< naywhayare>
can you send that to me too?
andrewmw94 has left #mlpack []
< sumedh_>
naywhayare: Hey ryan, I am really sorry... I went for a run... you still there??
sumedh_ has quit [Remote host closed the connection]
marcus_z1q is now known as marcus_zoq
sumedhghaisas has joined #mlpack
< sumedhghaisas>
naywhayare: hey ryan, any luck on that issue??
sumedhghaisas has quit [Ping timeout: 255 seconds]
imi has joined #mlpack
< imi>
hi
< marcus_zoq>
imi: Hello!
< imi>
:) I'm just watching the stanford machine learning course, and I was googling for C++ machine learning, that's how I found mlpack. Now, I don't really see how modules provided by mlpack will plug to each other. so for instance, let's suppose I have a binary classification problem. I don't see the functions that could be parameterized like anytype theta=learn(trainingset); neither bool l=getHypothesis(theta,someX);
< imi>
I see I can query K nearest neighbours, but if I do so, I still need to have an other container, a mapping from trainingCandidate to bool and query that then average the bool values I trained my container to get the hypothesis... now I think that's both a memory and a CPU overhead to look up twice (once from mlpack, then my mapping)
< marcus_zoq>
imi: In fact at this moment there isn't a k-nearest-neighbor classifier implemented. However, all required functions should be implemented, so maybe in the near future there is an implementation for this classifier.
< marcus_zoq>
imi: There are some other classifier like the naive bayes classifier, which implemented the requested interface.
< marcus_zoq>
imi: If you're interested you can open a ticket with a feature request.
< imi>
well... bayes classifier is based on bayes theorem, which is about conditional probabilities, but that's how far my knowledge goes in this exact moment... anyways I'd like to have a hypothesis mapping from int[60] or double[60] to bool (having a double in interval [0..1] is even better in some ways)
< imi>
(I mean having a a double in interval [0..1] instead of the bool is even better in ...)
< marcus_zoq>
imi: So if your just dealing with binary classification problems, I guess you can use logistic regression or a simple perceptron. Is this helpful?
< imi>
yes. thanks
< imi>
can I use that logistic regression stuff not to output either 0 or 1 but to return a double/float in range [0..1] ?
sumedhghaisas has joined #mlpack
sumedhghaisas has quit [Client Quit]
sumedhghaisas has joined #mlpack
< imi>
does the regression algorithms use gradient descent or Newtons method?
< sumedhghaisas>
imi: Hello imi, which regression algorithm are you referring to?
< imi>
all of them. they might use the same parameter (theta) optimalization algorithm
< imi>
if that algorithm cannot be plugged/changed using a template parameter as well
< imi>
can I use that logistic regression stuff not to output either 0 or 1 but to return a double/float in range [0..1] ?
< sumedhghaisas>
imi: if my memory servers correct linear regression uses simple gradient descent...
< sumedhghaisas>
*serves
< sumedhghaisas>
and logistic regression is templatized I guess.. the default optimization will be LBFGS... which is somewhat optimized form of newton raphson method...
sumedh_ has joined #mlpack
< sumedh_>
if you want to get a output in the range of [0..1] you require fuzzy techniques...
< sumedh_>
imi: my msges gettin to you?? my internet connection is bad nowadays...
< sumedh_>
ini: anyways... one way is to use the sigmoid values generated ... is that what you are asking?
sumedhghaisas has quit [Ping timeout: 255 seconds]
sumedh_ has quit [Ping timeout: 255 seconds]
< imi>
received
< imi>
newton method is quadratic in itself, so in my opinion there's no rational point in further optimizing