07:51 < jenkins-mlpack2> Yippee, build fixed!

07:51 < jenkins-mlpack2> Project docker mlpack nightly build build #281: FIXED in 3 hr 37 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/281/

08:26 < rajiv_> zoq: I had submitted my proposal through the official site last week, it would be great if you could review and give some feedback :)

08:30 < rajiv_> I'll also share the direct link here: https://docs.google.com/document/d/16pOrysmlU0nWvA8sbU_IEiemKd2ZlSvzTZySDwcR6cA/edit?usp=sharing

11:48 < Soonmok> Is there a way to get feedback on my proposals from MLpack contributors?. I've shared my GSOC proposals, and I wonder if I have to just wait or email to someone.

12:24 < zoq> Soonmok: Hello, you have to wait, we will give feedback once we have a chance.

12:25 < Soonmok> Thank you! I will wait.

12:25 < zoq> rajiv_: Already glanced over the proposal, but haven't left any comments yet.

13:19 < gauravcr7rm> zoq: Hey Marcus, I have already uploaded proposal on my dashboard and I also have given access to you to comment on it as asked by you. It would be great if you can give some feedback it. Thanks in advance : )

17:01 < robb9> how do i specify my own error function for a NN? Like my own MSE?

17:06 < zoq> robb9: as long as it follows the interface, see https://github.com/mlpack/mlpack/tree/master/src/mlpack/methods/ann/loss_functions for examples; you can pass it at instantiation time e.g. FFN<CustomLossFunction> model(..)

17:07 < robb9> where are those implemented: the cpp files?

17:07 < robb9> by that I mean where are they located

17:08 < zoq> mlpack/src/mlpack/methods/ann/loss_functions/

17:08 < zoq> mean_squared_error.hpp and mean_squared_error_impl.hpp are the files for the MSE loss function

17:22 < robb9> oh. They were both .hpp so I thought they were just header files

17:22 < robb9> thanks

17:27 < robb9> What is the Backward() function for?

17:28 < robb9> what if I just want to define a "fitness" function for how far it got on a certain task ? What would I make Backward() do?

17:30 < zoq> you have to return an error, one idea might be to take the maximum goal minus the achivement.

17:32 < robb9> I thought Forward() determined the error

17:33 < zoq> true, in your setting this would be the same

17:35 < robb9> ah, so I can just return the same output for both

17:35 < robb9> I'm probably going to do (1/"fitness")

17:35 < robb9> so higher fitness means lower error

17:35 < robb9> because technically there's no maximum fitness value

17:38 < robb9> zoq: thanks

18:12 < robb9> sorry, I have another question. I'm assuming Backward() helps the network learn, so how would backprop work if the output (error) is the same for Forward() and Backward?

18:53 < robb9> do the ensmallen optimizers use the error function I provided automatically or do I have to explicitly state it

18:57 < krgopal> HELP

19:43 < robb9> What if I want to evaluate the overall network, and I can't do it incrementally like Evaluate() does?

19:44 < robb9> because using my fitness function I cannot return the "error" of a specified single datapoint

21:00 < ReemGody> hello every one . I just want to apologize because I wa having problems with my mailing application as it required activation, so when I activated it, it sent all the mails I have written, so you may find the same mail sent from me multiple times on the mailing list

21:00 < ReemGody> Really sorry

21:03 < ReemGody> I have a question. I was exploring the tutorials on the library's repo and I found out that there were no tutorials for using reinforcement learning api, so I thought may be I could add a tutorial for this as I learn about the functionality.

21:03 < ReemGody> I could submit a pull request with that if you like the idea

21:28 < zoq> ReemGody: Hello, if you like to write something up, please feel free.

21:29 < zoq> robb9: ensmallen will only see the funcition you pass, and depending on the optimizer it calls Evaluate(...) which returns the loss and Gradient(...)

21:31 < zoq> robb9: If you batchSize = Number of samples, it coudl work

21:36 < zoq> robb9: The backward should return the error w.r.t the current input: I think if you take a look at the MSE layer it might be comes clear.

21:36 < zoq> Really like the new stackoverflow design :)

21:57 < rcurtin> the sparkles help me concentrate on the answer :)