If training fails due to hardware aging, what if we simply zero out the values coming from the affected device the moment the failure occurs?