09:54 < alexsc> hey there. I’m trying using NCA and I would be interested in knowing 2 things… 1. how to interpret the output matrix “distance”, that is not symmetrical: in order to transform a vector “x” to the new space, shall I do x ° A or A ° x ? 2. Is there a way to achieve dimensionality reduction with this NCA implementation? would be cool to use it, for example, for 2D visualization

09:55 < alexsc> with A I meant the “distance” matrix. A is the name in the original paper

11:29 < alexsc> after a few more trial and error, I’d say: I haven’t really understood how to use it… now checking the test for the NCA class I see that the matrix that’s output from LearnDistance can be fed into a Gradient object… still I don’t get how to find a projection matrix for my dataset that minimizes the KNN error, using NCA

15:53 < rcurtin> alexsc: sorry for the slow response---in order to use the output matrix from NCA, it depends on whether you are working in C++ where the data is column-major, or if you have used the command-line program or Python binding

15:54 < rcurtin> in any case, your matrix A will be non-symmetric and d x d, and your data matrix will be either d x n or n x d (again depending on what language you are using), so there will be only one way to multiply it

15:54 < rcurtin> for dimensionality reduction, I don't think you can do that directly with NCA, but, you could do a two-step process where you first stretch your dataset with the output of NCA

15:54 < rcurtin> then use PCA to remove the dimensions with minimum variance

17:25 < alexscc> back, if anybody answered please paste back… I’ll keep the connection alive now.. thank you

17:27 < alexscc> an example:

17:27 < alexscc> https://github.com/mlpack/mlpack/blob/6d4574d039b75e5bfe7bc85ca30df007219c41ca/src/mlpack/tests/nca_test.cpp#L283

17:28 < alexscc> I’d like to know how to apply “outputMatrix” to “data” in order to make it perform better with KNN, basically how to use NCA in this test

17:57 < rcurtin> alexscc: I answered, see http://www.mlpack.org/irc/

17:57 < rcurtin> I hope it helps

17:58 < alexscc> thank you!

17:59 < alexscc> yes, I get it. also the PCA part makes sense. It’s my first day with this framework and I am having a bit of a hard time.. I’m using C++

17:59 < alexscc> I’m converting from my current representation using

18:00 < alexscc> arma::mat data(database->dimensions, database->size());; // feature vector in the columns

18:00 < alexscc> I managed to run NCA with >> NCA<LMetric<2> > nca(data, labels); // labels are one rowvector

18:01 < alexscc> my feature vector dimensions is 8, now it’s crashing on

18:01 < alexscc> SoftmaxErrorFunction<SquaredEuclideanDistance> sef(data, labels);

18:01 < alexscc> double initObj = sef.Evaluate(arma::eye<arma::mat>(database->dimensions, database->dimensions));

18:02 < alexscc> weird though… it seemed to work before.. the objective went from -405 to -2.. I’ll keep debugging this tomorrow

18:03 < rcurtin> sorry I am out right now so I can't look into this too deeply, but when I get back home I can provide a better response

18:03 < alexscc> thanks for the help, no rush

18:05 < alexscc> I’ll go offline later and check the chat log. will be back tomorrow

18:08 < rcurtin> sounds good