arch1.m

独立主成分分析的工具箱

% Finds ICA representation of train and test images under Architecture I, 
% described in Bartlett & Sejnowski (1997, 1998), and Bartlett, Movellan & 
% Sejnowski (2002):  In Architecture I, we load N principal component eigenvectors 
% into rows of x, and then run ICA on x.
%
% Original aligned training images are in the rows of C, one image per row.  
% In the following examples, there are 500 images of aligned faces of size 
% 60x60 pixels, so C is 500x3600. This script also calls the matrix of PCA 
% eigenvectors organized in the columns of V (3600x499). If you use the pca 
% function in my companion code, then [V,R,E] = pcabigFn(C');
%
% The ICA representation will be in the rows of F (called B in Bartlett, Movellan & 
% Sejnowski, 2002): 

D = zeroMn(C')'; % D is 500x3600 and D = C-ones(500,1)*mean(C);
R = D*V; 	 % R is 500x499 and contains the PCA coefficients;

% We choose to use the first 200 eigenvectors. 
% (If PCA generalizes better by dropping first few eigenvectors, ICA will too).

x = V(:,1:200)'; 		% x is 200x3600
runica 				% calculates wz, w and uu. The matrix x gets 
				% overwritten by a sphered version of x. 
F = R(:,1:200) * inv(w*wz); 	% F is 500x200 and each row contains the 
			    	% ICA1 rep of an image

% Representations of test images under architecture I: 
% Put original aligned test images in rows of Ctest. 

Dtest = zeroMn(Ctest')'; % For proper testing, subtract the mean of the 
			 % training images not the test images: 
			 % Dtest = Ctest-ones(500,1)*mean(C);
Rtest = Dtest*V;
Ftest = Rtest(:,1:200) * inv(w*wz);

% Test nearest neighbor using cosine, not euclidean distance, as similarity measure.
% See my companion code for functions that calculate nearest neighbor using cosines.