1 | % Copyright 1996 Tony Bell |
---|
2 | % This may be copied for personal or academic use. |
---|
3 | % For commercial use, please contact Tony Bell |
---|
4 | % (tony@salk.edu) for a commercial license. |
---|
5 | |
---|
6 | |
---|
7 | % Script to run ICA on a matrix of images. Original by Tony Bell. |
---|
8 | % Modified by Marian Stewart Bartlett. |
---|
9 | |
---|
10 | %Assumes image gravalues are in rows of x. Note x gets overwritten. |
---|
11 | %Will find N independent components, where N is the number of images. |
---|
12 | |
---|
13 | %There must be at least 5 times as many examples (cols of x) as the |
---|
14 | %dimension of the data (rows of x). |
---|
15 | |
---|
16 | N=size(x,1); P=size(x,2); M=N; %M is dimension of the ICA output |
---|
17 | spherex; % remove first and second order stats from x |
---|
18 | xx=inv(wz)*x; % xx thus holds orig. data, w. mean extracted. |
---|
19 | |
---|
20 | %******** setup various variables |
---|
21 | w=eye(N); count=0; perm=randperm(P); sweep=0; Id=eye(M); |
---|
22 | oldw=w; olddelta=ones(1,N*M); angle=1000; change=1000; |
---|
23 | |
---|
24 | %******** Train. outputs a report every F presentations. |
---|
25 | % Watch "change" get small as it converges. Try annealing learning |
---|
26 | % rate, L, downwards to 0.0001 towards end. |
---|
27 | % For large numbers of rows in x (e.g. 200), you need to use a low |
---|
28 | % learning rate (I used 0.0005). Reduce if the output blows |
---|
29 | % up and becomes NAN. If you have fewer rows, use 0.001 or larger. |
---|
30 | |
---|
31 | B=50; L=0.0005; F=5000; parfor I=1:1000, sep96; end; |
---|
32 | B=50; L=0.0003; F=5000; parfor I=1:200, sep96; end; |
---|
33 | B=50; L=0.0002; F=5000; parfor I=1:200, sep96; end; |
---|
34 | B=50; L=0.0001; F=5000; parfor I=1:200, sep96; end; |
---|
35 | |
---|
36 | %******** |
---|
37 | uu=w*wz*xx; % make separated output signals. |
---|
38 | cov(uu') % check output covariance. Should approximate 3.5*I. |
---|