> For scientific purposes that same glitch could be catastrophic.

But for machine learning, some say that stochasticity improves convergence times!