It's important to note that what is being investigated are neuromorphic processors. i.e. attempts to model biological brains. A paragraph from the article direc

It's important to note that what is being investigated are neuromorphic processors. i.e. attempts to model biological brains. A paragraph from the article directly addresses why conventional regularization isn't used. > “The issue of how to keep learning systems from becoming unstable really only arises when attempting to utilize biologically realistic, spiking neuromorphic processors or when trying to understand biology itself,” said Los Alamos computer scientist and study coauthor Garrett Kenyon. “The vast majority of machine learning, deep learning, and AI researchers never encounter this issue because in the very artificial systems they study they have the luxury of performing global mathematical operations that have the effect of regulating the overall dynamical gain of the system.”