I wonder. Could the weight matrices that corresponds to the mean filtered training samples be good starting values?
I mean. It should be easier to correct a ?generalized starting image than to begin from scratch with random noise.
So my idea is to calculate and train on mean summarized collections of the input samples. Here I will use an extra label. Then use those converged weights as starting values. For when you want to generate.
So when I want to generate an image from label 2. I will try with the weight matrices from the mean calculation. Then let the algorithm work itself to an image that looks like label 2 from the extra label. That is from the mean filtered solution image to the image that looks like coming from label 2.