00:41 < zoq> rcurtin: Will open an issue.

01:18 < rcurtin> sounds good

01:18 < rcurtin> let me check if I can reproduce it first

01:18 < rcurtin> s/first/also/

01:20 < rcurtin> interesting, it doesn't happen with every dataset, but it does with optdigits

05:01 < jenkins-mlpack2> Project docker mlpack weekly build build #37: STILL UNSTABLE in 6 hr 14 min: http://ci.mlpack.org/job/docker%20mlpack%20weekly%20build/37/

06:17 < rajiv_> zoq: After I added Dense<> in layer_types.hpp in using LayerTypes = boost::variant<...>, I still get a similiar error: https://www.dropbox.com/s/1fza02w3j1024rv/build_log.txt?dl=0 I used arma::cube instead of arma::mat in Dense class as I had to concatenate for which I used arma::join_slices. Could that be an issue?

08:33 < ShikharJ> rcurtin: I think they're the standard practice for any C++ based project. Plus it would double up as a good first issue.

08:44 < jenkins-mlpack2> Yippee, build fixed!

08:44 < jenkins-mlpack2> Project docker mlpack nightly build build #256: FIXED in 3 hr 30 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/256/

09:12 < zoq> ShikharJ: Personally, I think the C++ cast style is somewhat bloated, especially for numeric casts.

10:28 < anuragsarkar250> join

10:29 < anuragsarkar250> hello everyone Im Anurag I was exploring different project ideas of MLPACK, would love to discuss more about it.

11:23 < siddhant> hii every one. I woud like to work for the project mlpack to tensorflow translator

13:30 < mayuri> Hi

13:30 < mayuri> I wanted to know how to contact the mentor or contribute to the project in which I am interested?

13:34 < Nisarg> I read a few research papers on constrained Particle Swarm Optimisation and also for multi objective problem . I want to send a proposal for this project. As mentioned in the ideas list, a good Gsoc proposal should contain possible places where the code has to be changed. Please guide me on what could be the best way to go for it. And also could you please send the format for the Gsoc proposal to the email I'd nisargbipinshah171me15

13:43 < favre49> zoq: I noticed in a mail that you mentioned that NEAT would havt to work with arbitrary functions. However, in most of the "learning to learn" papers you pointed me towards, they take a history of gradients as an input, or use CPPNs. How would we make NEAT work with arbitrary functions?

13:44 < favre49> The only way i can think of right now is if you were looking for just a direct mapping from input to ouput through NEAT, that couldnt be used for a different objetive or a different starting point

14:18 < mayuri> Hi

14:18 < mayuri> I want to know how to contribute to the project idea in which I am interested

14:19 < mayuri> If anyone could please tell!

15:36 < zoq> favre49: You are right the arbitrary function type didn't make sense in this case, but I think in this case you would have to write another optimizer that wraps NEAT, since the usage is somewhat different.

15:37 < zoq> mayuri: Hello, have you seen http://mlpack.org/gsoc.html and http://mlpack.org/involved.html?

15:38 < zoq> Nisarg: Hello, "For an application template, see the Application Guide." -> https://github.com/mlpack/mlpack/wiki/Google-Summer-of-Code-Application-Guide

15:39 < zoq> anuragsarka: Please feel free to ask questions here or over the mailinglist.

15:40 < zoq> rajiv_: How does the type you added to boost:variant look like?

15:41 < zoq> siddhant: Feel free to send an application.

15:52 < favre49> zoq: Thanks for the clarification :) I'll start looking for something applicable

15:55 < zoq> favre49: As for now, be could focus on arbitrary functions.

16:02 < favre49_> zoq: I understand this is in a way spoonfeeding, but are there any papers you can point me in the direction of that do this sort of thing? I've been struggling to find any. Even NEMO has no implementations as of yet, so it seems risky to implement without having an idea of its real performance.

17:43 < riaash04> Hi, I went through Bang Liu's implementation of the Neat (not very deeply just to understand the flow and major functions), and It seems very good. Although, I was wondering if it's expected to build on that (like Kartik did) or a complete new implementation could be done for this year's GSOC idea?

17:47 < riaash04> Also, in the proposal should I mention my pr(if it gets merged), even though it's not directly related to this project ?

18:40 < rcurtin> favre49_: I think a lot of the optimizers that we implement we don't actually have a guarantee that they will work well

18:40 < rcurtin> here's a fun fact: the vast, vast majority of papers introducing new optimizer techniques compare against other optimizers using only a *single* random initialization

18:40 < rcurtin> therefore even though their optimizer may converge faster for *that particular starting point* they use in their experiments, there's no understanding of whether it will across lots of different problems

18:41 < rcurtin> so I think a lot of the things we implement, we can't totally be sure if they are good, and we have to do some comparisons and simulations afterwards :)

18:54 < shadowfox> is knowing ml a pre requsite for your projects?

18:57 < shadowfox> ?

18:57 < shadowfox> there?