The nice thing about this particular application (deep neural networks) is they're usually fairly insensitive to weight noise because they have a lot of redunda

The nice thing about this particular application (deep neural networks) is they're usually fairly insensitive to weight noise because they have a lot of redundancy, just like our brains. Your question still stands, however. And for an analog chip it's basically the question of how to qualify, validate, and characterize neural networks in general. I'm not convinced we have the faintest idea how to do that in any formal sense.