Show both heads all data or use auxillary overclustering?

[vivektreddy]

In the case of STL10, it says that using overclustering with an extra 100000 images helps improve clustering of those initial 5000 images significantly.
What would have been the trade off if you had ran the algorithm with all 105000 images and showed each head the same images. Is there a better clustering achieved when you use auxillary head performing overclustering.
If I have 200,000 images for fully unsupervised clustering, would it be better to show both head A and head B all 200,000 images? Or would it better to shown one head 20,000 and the other head 200,000 (with more clusters)?
Furthermore, if I show both heads the same images, I might as well use one head right?

Thank you.