How Do You Preserve Your Performance With Network Pruning?

Doyou: Pruning seems to only work after training, or at least that was the conception. My result said, actually, it may be a litle more complicated than that. You can find sub networks that are just as small as you would get from pruning after training. They existed at the beginning of training, you can train them to completion and get the same accuracy. But i don't know how to find these networks at the beginning. And this is a research question i hope my work has inspired people to ask. Doyou want to talk more about the paper? It'd be interesting to hear, like, you know, how have your views grown on it,