GMM preprocessing steps

2017-09-07 07:46:53

I am trying to classify MNIST data (28x28x60000) using a GMM. I am looking for hints on preprocessing steps, what I've tried so far:

PCA and run GMM on first 10 principal components, looking for 10 GMM components: very poor performance

ICA and run GMM, slightly better performance than PCA but still not great

i've tried some manifold dimensionality reudction techniques like isomap and MDS but these don't work on such a large dataset, at least not on my pc.

In both cases I have tried initialising GMM weights with means equal to Kmeans cluster centers on the PCs.

bonus question: if using GMM with more than 10 components, is there a way to map these components back to the original 10 classes?