The Importance of a One-to-One Correspondence | 2min snip from Machine Learning Street Talk (MLST)

Speaker 1
of the fact that it exists. Right,
Speaker 2
right.
Speaker 1
So
Speaker 2
yeah, I think there's an interesting point in there, which is that it isn't this exact one-to correspondence. And I mean, really, how could it be, right? Like how could a, how could say a subset of the system precisely represent the entire system, but instead it, in a sense, it's representing an abstraction. Absolutely. You know, and I mean, I think there are even good energetic arguments why that would be the case, because, you know, you, you can't maintain total information, you maintain an abstraction of it, you maintain enough in order that you can predict and track, you know, behaviors. Because you exist, you must continue to exist. But you don't require complete information or sufficient
Speaker 1
tracking. Absolutely. Yeah, heuristically, you can think of the free energy principle, and this is metaphorical, of course, but as a map of that part of the territory that behaves like a map. So it is a scientific principle that we can use to construct models of systems that appear as if they are in turn modeling or tracking aspects of their environment. And I think thinking about modeling for a second is useful here. So if you had a one-to scale map, is it Borges who presents that in a story at some point? Yeah, I think so. Yeah, so a one-to scale map would be completely useless, right? I mean, just imagine the quantity of paper that you would need to have a map of, say LA, right? A one-to scale map of LA. It wouldn't be useful in the generation of your actions. So the map is necessarily simpler than the territory, but that's okay in some sense, right? Because the implication of that is that using the map, you're going to encounter errors. You're going to generate errors. But in the active inference and the free energy principle approach, those errors are the relevant signal, basically. informed by what your map contains in terms of information, and to in real time course correct based on the errors that you're generating. So these errors and the kind of oversimplified character of the model are features rather than bugs. You would need that in order to have a signal at all in some sense.

Dr. MAXWELL RAMSTEAD - The Physics of Survival

Machine Learning Street Talk (MLST)

Remember Everything You Learn from Podcasts