I could see a really good prestige mechanic for this being improvements to the learning algorithm. In particular, I could see updates to how the prior model is calculated to be an ensemble of the most recent run and others. I could also see updating the algorithm parameters to change how much randomness is introduced with each iteration.
 "Idle Machine Learning"){kind=icon}