08:15 < marcus_zoq> Udit_s: Hello, Is it possible that you forgot to verify the results in the second, ..., fifth test?

08:20 < udit_s> marcus_zoq: Hi !; hold on, - yeah I think I did. I was banking on going through the csv files to verify them; I think I planned to change them later on. I'll get that fixed.

08:20 < marcus_zoq> udit_s: Ah, okay.

08:21 < udit_s> Also, I was just going to get started on the perceptron.

08:21 < marcus_zoq> udit_s: Okay, great.

08:23 < udit_s> I wanted to talk about the learning policies - I wanted to properly understand at least the pocket policy. And I was wondering if you could share a few links which properly explain how the others - BEAM, PRISM work.

08:23 < udit_s> Right now, I'm just going to implement a basic multi-class perceptron. Then, I'll be extending it as we talked about, through template parameters.

08:28 < marcus_zoq> Sounds like a good plan, I'll see if I can find some good explanations. The pocket algorithm should be pretty straightforward. I think we have agreed to implement only one method, right?

08:31 < udit_s> I still wanted to think about implementing another one.

09:24 < marcus_zoq> udit_s: Here is a nice paper about a learning algorithm that is called Random Coordinate Descent:

09:24 < marcus_zoq> http://www.work.caltech.edu/~ling/pub/tr05perceptron.pdf

09:24 < marcus_zoq> udit_s: They compare the random coordinate descent algorithm with stochastic gradient descent, pocket and averaged-perceptron. The pseudo code is also quite instructive. Right now I'm looking for the paper that combined beam with the pocket algorithm...

10:37 < udit_s> Great ! Thanks.

11:43 < naywhayare> udit_s: I am here too for the next couple hours; I'm going to try to go through the decision stump code now and in the next handful of days

11:44 < naywhayare> technology is crazy! I'm sitting in a bus going down the highway in rural Georgia and I'm connected to the internet. couldn't do this ten years ago...

13:45 < Anand_> Marcus : You changed some code? Were we doing it wrong?

13:48 < marcus_zoq> Anand_: Yeah, I've made a few changes to the matlab code. Instead of 'NaiveBayes.fit(TrainData, 'Prior');' we have to use 'posterior(classifier, TestData)', I'm sorry for the misinformation.

13:48 < marcus_zoq> Anand_: If you like you can test the current version on the build server, but I've tested the code and it works.

13:48 < Anand_> Great! No problem! :)

13:50 < Anand_> We need to take shogun next, right?

13:52 < marcus_zoq> Anand_: If you like we can modify the code :)

13:52 < Anand_> yeah sure, we will

13:52 < Anand_> Pointers?

13:53 < Anand_> And btw, we need to merge with the master. Is it the right time? Maybe tomorrow?

13:58 < marcus_zoq> Yeah, tomorrow sounds good. Maybe it would be good to ask (IRC) the shogun guys whether there is no option to get the probabilities.

13:58 < marcus_zoq> Anand_: Maybe there is a hidden feature...

13:58 < Anand_> I will look at the code and see if I can figure out. Also, you meant on #shogun ?

13:59 < marcus_zoq> Anand_: Yeah, #shogun!

13:59 < Anand_> Ok, I will ask someone there. Should I mention gsoc with mlpack?

14:02 < Anand_> I will get back to you

14:06 < marcus_zoq> Anand_: There are all cool guys. I think you can just ask if there is an option in the nbc method to get the probabilities :)

14:06 < marcus_zoq> Anand_: Btw here is the code: https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/GaussianNaiveBayes.cpp

14:08 < marcus_zoq> Anand_: This the the parameter that stores the propb: 'SGVector<float64_t> m_label_prob;'.

14:08 < marcus_zoq> Anand_: We can use a 'nice' warkaround to get the vector :)

15:14 < Anand_> Marcus : So far no one seems to be responding on shogun. I am waiting

16:44 < marcus_zoq> Anand_: Okay, if you look into the code there is a protected parameter that is called m_label_prob. I think we can easily add a fuction that returns this parameter. Or use a workaround something like: '#define private protected' but I would prefer the first way.

17:07 < Anand_> Marcus : Ok. I will look into it if we can use that