.The Physics of Survival - Machine Learning Street Talk
Maxwell Ramstead "[...] what's transpired over the last few years is that really you can think of the free energy principle as a kind of generalization of the second law of thermodynamics to open systems. So, you know, the kind of universality that the second law has with respect to closed systems, the free energy principle has with respect to systems that are far from equilibrium."

"[...] in the case of the free energy principle, if you survive, if you exist, you're inexorably drawn to the set of dynamics, to the set of mechanics. Otherwise you don't exist. Well, to get technical for a second, I think there are two issues that are both striking and that speak to what you just said. So things like the principle of least action and the free energy principle and the principle of maximum entropy, they are in some sense true a priori or mathematically they are mathematical truths. So you wouldn't try to falsify the principle of least action empirically any more than you would say, try to falsify calculus or probability theory by coming up with an empirical counterexample. So there's a sense in which, like the the truth of these statements is robust and mathematical. Having said that, it is a striking empirical fact that the physical universe does in fact seem to conform to these mathematical regularities."


.Autopoietic Enactivism and the Free Energy Principle - Prof. Friston, Prof Buckley, Dr. Ramstead
Maxwell Ramstead "There's this sense I think in the literature that there are, like, 17 different possible interpretations of the free energy principle. And that you can make me say 1 thing and its opposite. What I've been going around saying is, no, that's not correct. There's only the free energy principle and it has the implications that it does, and there are misinterpretations of the free energy principle." 


.Autopoietic Enactivism and the Free Energy Principle - Prof. Friston, Prof Buckley, Dr. Ramstead
Maxwell Ramstead "I think the, kinda core idea to remember is that, minimizing [...] your variational free energy with respect to a generative model (right? So the usual free energy story), is dual to or equivalent to maximizing your entropy with respect to a set of constraints. So when the enactivists talk about constraint closure and all of this stuff, you know, Alicia Gerrero's new book, context changes everything (the enactivists in the autopoietic tradition really want to go hard on this and say, well, what you really need to do is map a set of constraints such that the system kind of self generates), well, that is rigorously and absolutely equivalent to minimizing your free energy with respect to a generative model. There's no difference there. So we're talking about the same thing. We're talking about the way that, like, a set of structured dependencies generates a thing that sustain itself over time. And, as I was saying, this 'sustained over time' is a way of describing the interface or coupling between the system and its environment. So, that, just mathematically we're talking about the same thing."

Keith Duggar's answer "And they seem to get upset when any attempt is made to link organization to structure. For example, you know, to me, the markov boundary that say, analogous to an animal cell membrane, I understand, as a former person who studied biology, that the molecules of that membrane are in flux, things come and go, proteins are added, carbohydrate rates are removed, etcetera. I don't have in my mind a set of molecules that's like they're forever in some type of like, you know, stationary structure. It's a very dynamic kind of system."


.On the Map-Territory Fallacy Fallacy
"This paper presents a meta-theory of the usage of the free energy principle (FEP) and examines its scope in the modelling of physical systems. We consider the so-called `map-territory fallacy' and the fallacious reification of model properties. By showing that the FEP is a consistent, physics-inspired theory of inferences of inferences, we disprove the assertion that the map-territory fallacy contradicts the principled usage of the FEP. As such, we argue that deploying the map-territory fallacy to criticise the use of the FEP and Bayesian mechanics itself constitutes a fallacy"

.The Physics of Survival - Machine Learning Street Talk
Keith Duggar "Usually we think, 'what does a lifeform, or a thing even for that matter, has to do to survive?'. The emphasis of 'are you fit', the darwinian sense of what kind of fitness does it take to survive? But the free energy principle completely inverts that, and it says: 'okay, if things exist, if things survive, what must they do?' Right, and it turns it on the head in this way, which is : let's assume that there is a thing, you know, and it continues to survive. It continues to exist. Just by knowing that, what must it do? What dynamics, what behavior must it have? Is that a fair framing? And what is, what are those behaviors that 'things that exist' must do?"

Maxwell Ramstead's answer "I think the way that you describe things is accurate and an insightful way of putting things. The free energy principle is not just a hum... Basically a theory according to which things that exist must do this and that. As in it's not trying to tell you 'here is something that things do in order to exist'. What it's telling you is 'we observe that things exist, in the sense that there are systems or particles or things that can be reliably re-identified, that are separated from but coupled to their environment. And given that we observe these things that exist : 'what must be true to them?' So it's a kind of inversion of the explanation, moving to like a kind of first principles account of what must necessarily be the case if you exist. And essentially what the free energy principle tells you is that if you exist, in the sense that you are separate from but coupled to an embedding environment, then it will look as if you're tracking the statistical structure of your environment or, more precisely, it will look as if the states and paths that are internal to your boundary (the boundary of a given thing), track things that are external to that boundary. So in some sense it explains why hum... Or it provides a principle allowing us to explain why it look as if everything that exists is 'tracking' or 'representing' depending on how you think about it, features that are external to it. And this tracking or representing relation, it is rather 'weak' in some sense, but we're not talking about like necessarily contentful representations (little images in my head). What we are talking about is something I think more fundamental or existential."

.The Physics of Survival - Machine Learning Street Talk
Maxwell Ramstead "Why use active inference : it is demonstrably the most efficient machine learning technique." 

"What the free energy principle allows us to formalize is the thermodynamics of information writing onto the boundary. [...] So in some of the newer work on the quantum information theoretic formulation of the free energy principle, which we don't necessarily have to get into in detail, but there are these kind of new scale free extensions to the free energy principle that have been developed that appeal to the tools that have been developed in quantum mechanics, right? [...] So the kind of information theory that gets augmented to handle things like probability amplitudes, which are the the roots of probability densities. And so you can get your wave equations moving. Place and all that. So the that formulation of the free energy principle allows us to formulate the computations carried out by a system in terms of like a per bit read and write cost. So there's a sense in which like you're bringing it down to the like to the bare kind of, you know, machine elements of your computations and you're writing things down in a way that is demonstrably the most efficient way of doing it."


.Designing explainable artificial intelligence with active inference: A framework for transparent introspection and decision-making

.A free energy principle for generic quantum systems


.The Physics of Survival - Machine Learning Street Talk
Keith Duggar "I think there is an interesting point in there which is that it isn't this exact 1 to 1 correspondance. I mean really, how could it be? Like, how could say a subset of a system precisely represent the entire system, but instead in a sense it's representing an abstraction of the system." 


Maxwell Ramstead "All you have to do is have a good enough map and act in such way that you are informed by what your map contains in terms of information, and you in real time course-correct based on the errors that you are generating. These errors in the kind of oversimplified character of the model are features rather than bugs. You would need that to have a signal at all in some sense."

Keith Duggar "Yeah. And there is this iterative nature to it that I think is sometimes forgotten because you know, there are these two components in the free energy principle. One is fidelity. How accurately does it kind of map to the environment? So the entity that's surviving, the thing that's surviving, you know, has to have a model of the environment, has some degree of fidelity because if it doesn't, it's not accurate enough to maintain its survival. But at the same time, it also has to have adaptability, right, because the information is never complete. There's always new, phenomenon occurring to it, the environment is changing or whatever. So the model has to maintain a degree of flexibility. And that's what this kind of entropy term in [FEP] is. It's saying 'you need to maintain a certain amount of entropy because that is a form of flexibility'. Is that correct?"

Maxwell Ramstead's answer "I mean, that's absolutely correct. And you can think of the entropy in a few different ways and the entropy term in previous discussions with Karl Friston, you have highlighted its technical importance. I mean, basically what we're trying to do when we minimize free energy is to increase the predictive accuracy of our model. So that is to have a model that generates predictions that are as close as possible to the real data that I'm ingesting. But the free energy principle allows us to, in a principled manner, penalize the complexity of the model. Right? Because you don't just want an arbitrary explanation, as you know, you can construct an arbitrary explanation for any data set."

Keith Duggar "It can even be deleterious if you have incomplete information and you model it too accurately, accurately in a loose sense, then actually you're memorizing spurious information that doesn't generalize."

Maxwell Ramstead "Absolutely. Yeah, absolutely. And so the free energy principle, when you're applying it and you're saying that systems that exist minimize this quantity, variational free energy, the variational free energy can be decomposed roughly speaking, into predictive accuracy minus complexity. And so what you're doing is you're penalizing your gains in predictive accuracy against the complexity cost of your model, basically penalizing every new degree of freedom that you need to introduce into your model to explain the data. So in some sense, the free energy principle, you can think of it as kind of statistical predictive accuracy, but also Occam's razor" 

Keith Duggar "It's an interesting balance. And the free energy principle encodes that balance. And [...] this is a sense in which the free energy principle applies to itself because it's. That's right. It's almost the simplest formulation of that of that balance, right?"

.Weak Markov Blankets in High-Dimensional, Sparsely-Coupled Random Dynamical Systems

.Active Inference: The Free Energy Principle in Mind, Brain, and Behavior



.The Physics of Survival - Machine Learning Street Talk
Keith Duggar "So that's interesting. In a universe like ours that has the basic physics that a universe like ours has, as the scale of a system gets larger and larger, you generate Markov blankets..."

Maxwell Ramstead "You're bound to with a probability one." And you know, most of the systems that we consider in physics are large in the appropriate sense, right? So think about how many molecules are in a drop of water. It's [...] ten to the 23. That's just for a single drop of water. Now if you consider the brain, the brain has like something on the order of 100, 150 billion neurons, each of which make thousands of connections. If each of those connections can encode a parameter, then you're talking about like a very large system, right? We're way, way, way beyond like, you know, 20, 50, 1000 different states that are coupled together. We're talking about like billions and trillions of different states. So there's reason to think that just due to the physics of the situation, most relevant things that we might want to consider will have Markov blankets"

Keith Duggar "[...] So we have a measure of blanketness. It's kind of between 0 and 1 zero has a blanket, one doesn't. Okay. And and yet as the scale of the system gets larger and larger, blanketness approaches zero, you get blankets no matter what. And in a sense there's a sense in which that's recapitulating what we see. If we just look around like everybody out there listening, look around yourself and you're going to see blankets all over the place. You're going to see things. And those things have boundaries. But it's remarkable, right, that there's a mathematical proof that that's inevitable in this sense, isn't it? " 

Maxwell Ramstead "Well, I think it's remarkable in part because we have approached the question of self-organization and emergence from a false starting point. So I've been going around saying recently Aristotle was wrong. That's that's sort of my philosophical start. Well, the whole is much less than the sum of its parts, it turns out. So yeah, there there are a bunch of things to unpack from that. Well, the first is that what makes you the kind of thing that you are is the sparsity of your coupling to the rest of the world, right? If you think of a gas right where everything is coupled to everything else, then it's just this fuzz and it's all one system and there's there's no you can't really identify particles within the system. Particles are things are defined by their sparse connections to everything else. So I am in some sense what I am not or I can be defined in terms of what I'm not connected to as opposed to what I am connected to. I mean, if you were to create like a giant adjacency matrix for the entire universe, most of it would be empty, right?

[...] but there's more. Think of an engine. Like an engine functions as an organized whole because you're constraining its parts to behave in very specific ways. So like, you know, if you think of an engine more specifically like a petrol engine, well the mechanical effect of the engine you get by moving these pistons in a specific direction up and down. And the best way to wreck your engine is to introduce new degrees of freedom into it. Right? I would not want to introduce new degrees of freedom into the Pistons. That's a that's a great way to just tear your engine apart. And I would submit to you that this is, you know, an inaccurate way of thinking about all self-organization. We exist as wholes because our parts are constrained to behave in very specific ways.

So it's not merely that I am what I am because I am not what I am not. It's just a nice tautology. It's that what makes me what I am is the way that I remove degrees of freedom from my parts such that they conspire to create, you know, to generate me as an overall pattern."

New to LessWrong?

New Comment