Where does the variance come from?

[jakobnissen]

We know that the accuracy of Vamb differs quite a lot from run to run. When benchmarking, we run each setting three to five times in order to average out the run variability.
But where does this come from? Almost certainly, it comes from the randomness in the neural network training, or the randomness in the clustering algorithm.
It would be nice to get a breakdown of how much of the variance is explained by these two steps.