Model-Based Learning

G-Fluonets is one incarnation, but there's the bigger principle of how you model something compactly could be separate from your infant system. Because you model something separately more compactly, you need a lot less data. And that's the promise of being the promise of model-based RL since the beginning right? True. But we can apply this thinking not just to RL. This could be also in cases where there is no sequence of decisions. It's just there's knowledge and there's inference, questions you'd like to answer. So it's very plausible that we're like our inference machine trains itself on to be consistent with our knowledge.