Dictionary matrix for online sparse coding doesn't learn

Following Online Dictionary Learning for Sparse Coding and using an inference function I've found on the Stanford's website, I'm trying to online learn a Dictionary matrix. Tested both on resized actual images and some revised version of MNIST, I see that I do manage to get something to happen, but the results just doesn't look anything like real feature detectors. The way I train is by taking an image, lets say 500500, then go through all 1515 patches (1px progress each iteration), do the inference and the dicotionary optimization.

The main file:

for i = 1 : height-atomSize
    i
    for j = 1 : width-atomSize
        xt = imgG(i:(i+atomSize-1), j:(j+atomSize-1));

        xt = reshape(xt, pixNum, 1);

        h = TwIST(xt, D, 7);

        A = A + h * h';
        B = B + xt * h';

        D = dictUpdate(D, A, B);
    end
end

The dictUpdate function:

for i=1:size(D,2)
   D(:,i) = (1/A(i,i)) * (B(:,i) - D * A(:,i)) + D(:,i);
   % L2 norm:
   D(:,i) = D(:,i) / max(norm(D(:,i),2), 1);
end

Another thing, if i'm already asking here, I'd like to do the inference myself and have found a vectorised version for it, but every time I test it it simply explodes:

h = h - alpha * D' * (D * h - xt);
h = shrink(h, alpha*lambda);

And the Shrink:

for i=1:size(h,1)
    h(i) = sign(h(i)) * max(abs(h(i)) - alphaLambda, 0);
end

Thanks for the answers