probml / pmtk3

Probabilistic Modeling Toolkit for Matlab/Octave.
MIT License
1.55k stars 796 forks source link

fisherDiscrimVowelDemo (pca-part issue?) #62

Open ido opened 10 years ago

ido commented 10 years ago

From email.po...@gmail.com on September 02, 2013 08:47:39

What steps will reproduce the problem? 1.Run the fisherDiscrimVowelDemo.m What is the expected output? What do you see instead? In the "%PCA projection to 2d" part of the code the function "pcaPmtk" is invoked twice: (i) in line 12, for the whole dataset and (ii) in line 17, for the class means.

Wouldn't this correspond to two different projection subspaces? In other words, wouldn't B and Bmu span two different linear subspaces? But if this the case, shouldn't we avoid plotting the projected points and projected means in the same plot (see lines 41-48)?

Would it make more sense to invoke the function "pcaPmtk" only once (for the whole dataset as it is done in line 12) and then project the class means to the linear subspace spanned by the basis vectors B (already found using pcaPmtk)?

I would therefore replace line 17:

[Bmu, muC2d] = pcaPmtk(muC, 2);

with

muC2d = centerCols(muC)*B;

Does this make sense?

Any feedback would be greatly appreciated.

Many thanks What version / revision of the product are you using? On what operating system? pmtk3 / Windows Please provide any additional information below.

Original issue: http://code.google.com/p/pmtk3/issues/detail?id=62