One more query @vperetroukhin:
Regarding the PCA which is optionally used for initializing classification, am I understanding this correctly:
- Select random subsets of particles (many more subsets than desired classes)
- Reconstruct each subset
- Cluster the subset-reconstructions using PCA (similar to the approach used in 3D-VA display for clustering…?). Average “superclusters” to give desired number of classes.
If this is right, when is the initial lowpass filter applied? Is it applied right at the end to the averaged clusters, or initially, to the reconstruction of each subset?
Also, how many reconstructions should we use? the default is 100, but I guess that won’t be appropriate if you want 100 output classes… how many subset reconstructions approximately should one have per output class?
Cheers
Oli
