2,007 followers
@mckinley_scan @MichalSofka It's similar in this work https://t.co/50G8rcFBHy which may be what Michal was thinking about? Except they make each iteration into a network layer, so it can be trained end-to-end. But still each layer checks for consistency.