 All right, hello and welcome everyone. It's March 17th, 2022. We're here in Acton Flab, live stream number 40.1, discussing a free energy principle for generic quantum systems. Welcome to the Acton Flab. We're a participatory online lab that is communicating, learning and practicing applied active inference. This is recorded in an archived live stream, so please provide us with feedback so we can improve our work. All backgrounds and perspectives are welcome here, all QRF, and video etiquette for live stream will be hopefully following. To learn more about Acton Flab, head over to activeinference.org and for the rest of Acton Stream 40.1, blue, thanks for being the facilitator. Sure, so in this stream today, our goal is to learn and discuss this paper of free energy principle for generic quantum systems by Chris Fields, Carl Friston, James Blaisebrook and Michael Levin. And we have both Chris and Carl on the stream with us today. So if we, maybe we wanna just start off briefly with doing some introductions. So my name is Blue Knight. I'm an independent researcher in New Mexico and I will pass it off to the first author, Chris. Hi, good morning and thanks for attending this discussion. This is a paper about active inference and how to place the active inference framework in a quantum information theoretic setting. So what I thought I should do to introduce it is just give a little bit of the motivation for doing that. Um, quantum information theory is an outgrowth of quantum theory that started in the 70s or even you could take it back to the 60s if you wanted to. And its goal was to consider physical interactions as information exchange processes and to reformulate quantum theory in this way. And so it has a very natural kind of overlap with what Carl has been doing with the free energy principle in terms of reformulating physics, essentially all of physics and information and decision theoretic terms. So why quantum theory? Why go to this theory that's often thought of as a theory only of very tiny objects that are moving very fast? And the reason is that it's scale free and conceptually quantum theory is very simple. And it has one core hypothesis which is the principle of uniterity. And the principle of uniterity is simply the idea that total information is conserved. So quantum theory is based on a conservation law, a proposed conservation law, conservation of information. So what going to quantum theory really forces us to do is to take seriously the project that Carl spent a lot of time on in his particular physics paper in 2019, which is to understand what we mean when we say that we've observed a system multiple times. What we mean when we say we've interacted with the same object multiple times. And it forces us to, or at least it encourages us to take that question seriously and it provides a number of tools for talking about that question. And we'll talk about that as we get into the paper. But, and that's in a sense our motivation is to take this very nice, quite abstract, but in a sense very fundamental toolkit that's based around the idea of conservation of information and apply it to this question of active inference. So I'll let Carl pick it up from there. Right, well thank you. It's a great pleasure to be here. I have to say I was hoping to learn more than provide a background. So this is a great opportunity from my point of view is to hear a conversation of the Department of Painting in terms of how this program of control scale free background, that's what I think is one of the most attractive parsimonious benefits of taking a content perspective on this. And I think that the quinian desert landscape of minimal assumption, which I'm always drawn to and apply this to some fundamental questions which have actually been beautifully listed for us at the top here about being some crucially how things couple to other things in terms of measurement and observation and inference. So I'm personally delighted and hungry to learn about how simple questions about the nature of England we observe and measure and make inference about can be cast in this fundamental or from this fundamental perspective of quantum information theory. Very nice, thank you. James, thanks for joining us. Would you like to introduce yourself and maybe say a few words about the paper or pick it up from there? Oh, hello. You may call me Jim. Thank you. You want a potted biography, do you? Anyway, I'm a retired academic emeritus professor at Eastern Illinois University and adjunct associate professor at the University of Illinois at Urbana-Champagne. My background is global analytic geometry and some related areas. But in the last 16 or 17 years, I've been applying mathematical techniques to cognitive neuroscience and foundations of quantum biology and more recently quantum information as it relates to the free energy principle. Great, thank you. Stephen, do you want to maybe introduce yourself and give a thought about the paper as something you like or remembered or were curious about today? Yeah, thank you. Hi, I'm Stephen Sillett. I'm based in Toronto and I'm just doing a practice-based PhD into social topography and community development. I have got a background many years ago in chemistry and I'm really curious about the way that this idea of something that's scale-free also ties to something which seems to be very specific on scale. I wear something that's highly contextual and highly bounded by degrees of freedom at a certain scale and then how something could be inferred that's across all sorts of scales. That's really something that I'm curious about on this particular piece of work. Great, thanks. Daniel, do you have any other particular specific questions or... Just looking forward to seeing the discussion and anyone who's watching live can ask a question. So I'm curious about how, also like Stephen said, the specificity of a given experiment that interfaces with general aspects of the formalism and many other things. But we'll go ahead, Lou. Well, do we want to maybe start there? Do you guys want to speak to that scale-deterministic aspect of the quantum FPP? Well, I can say something here. Any information theory we'll start with some state space. So some set of degrees of freedom that have some set of values. And in the paper, we restrict ourselves to finite state spaces because we're really interested in systems that only have access to finite energy and so finite computational resources. So one starts with a state space and in a sense that state space when viewed from kind of the external perspective of the theorist who's defining it sets the scale of the problem. Because in defining the state space, you're saying what the values are of the degrees of freedom. And when you say, what are the possible values you are implicitly making some assumptions about resolution and making some assumptions about measurement. So for example, if we think of a canonical quantum system which is a qubit, just one, a particle with a spin and we say the spin can be up or down, then we've made an assumption about how we can measure the spin. And the most natural mathematical framework is a three-dimensional ball of values or a sphere of values if you normalize them. And that's making an assumption that we can measure the spin in any one of three directions. And so here we have imposed some possible values on this set of degrees of freedom. And so in a sense, we've set a scale for the problem. Similarly, if we're talking about interactions between human beings, clearly there are lots and lots of degrees of freedom that we can think of, but we typically think about only a tiny handful of degrees of freedom that have to do, for example, with verbal communication or visual communication or olfactory communication or some other kind of communication. And we always impose resolution restrictions. We can say this difference is detectable and these other differences aren't detectable. And whenever we make a decision like that, we're setting a scale. And when we characterize the information theory as scale-free, we're just saying that the formalism doesn't change when the scale changes. So the formalism, the mathematical description remains exactly the same, regardless of what scales we pick by choosing the possible values of the degrees of freedom that we're talking about. We don't change anything else about the way the theory works. So that's all that scale-free means. It doesn't mean that we've somehow abolished the notion of scale. That's part of our experience. So it's not something we're gonna make go away. It's that we haven't changed the way the math works depending on the notion of scale. And the most notorious example of that thing that we're not doing is the historical division of the world into kind of a quantum domain, a classical domain and a cosmological domain where you use three entirely different theories, right? Quantum theory, classical physics and general relativity to talk about what's going on. So that's not scale-free. There, you're radically altering the mathematical description to deal with different kinds of problems. So that's what we're not doing. Kind of use the same mathematical description to deal with all problems. I think that's a really helpful description, Chris, because so many times when we think about quantum, we think about the very, very, very small, like subatomic particles. And so it's pretty amazing to see math that's interchangeable with my level of existence, like not just tiny little pieces of me being used from quantum theory. What, is there any other math that you know of or any other examples of math that's been interchangeable with like classical and cosmological and quantum scales ever before, or is this maybe the first time that's ever been done? Well, no, if you look back to classical physics in the era of Newton, or even I think the better representative is classical physics is reformulated by Laplace. In that era, electromagnetism wasn't really understood yet. So people were thinking about classical physics in terms of mechanical interactions and gravity. But the idea at that point was that forces act instantaneously and forces act with one over R squared dependence. That was the only force that was really understood, the force of gravity. Forces were symmetrical, right? Newton's second law, action and reaction. And this theory was meant to be universal. And so if you look at the structure of that theory, as various people have pointed out, I think Nicholas Gison has written a couple of papers about this. It's a non-local theory. And even in Laplace's era, they realized that there was a deep issue with this theory that if a rock rolls downhill on the moon, then that immediately affects the state of everything on earth. In fact, immediately affects the state of everything in the universe. So that starts to look a lot like entanglement that states are not locally independent. And that's the fundamental thing that Einstein attacked about classical physics in formulating the theory of relativity was to insist on non-instantaneous forces, which just means forces transmitted by some carrier, which we call light. But light is really just a carrier of information. It could have been anything, it could have been neutrinos or something like that. So what Einstein did that was most revolutionary was to impose locality on physics. And that in a sense divided the world into realms that from a formal perspective, not just a practical perspective could be treated independently. And so it's no surprise that Einstein hated the idea of entanglement because entanglement was undoing precisely what he had done to, in his mind, repair physics. So yeah, I mean, classical physics of the time of applause was formally a scale-free theory. That's awesome. Thank you for the clarification there. So Einstein broke physics. Stephen, did you have a question? Yeah, just one question sort of following from that is, as someone sort of thinking about how this might play out in other ways, is it useful to think about things all the way up from the smallest and systems all the way up down from the biggest and classical has been when we were trying to name things as systems because often I hear people talking systems, systems, systems only but the use of things seems to be important as well. And I'm wondering if we should have more thinness in our way of looking at the world. Oh, should I comment on that? Yeah, okay. Yeah, I think you're raising your really interesting question. And when we talk about things as Carl so beautifully emphasized in his 2019 paper, we are implicitly talking about something that has an identity and that maintains that identity through time. And that's a deep philosophical assumption that's nearly always implicit. I mean, people just don't talk about that assumption very much. And when we use the word system, I think often trying to speak a lot more loosely but in fact, our language doesn't let us speak more loosely. If I talk about a system, then implicitly I'm saying I've got a particular state space. That's what a system is. And if I talk about a different state space then I'm talking about a different system. So let's apply that to something interesting like embryology. I pick a set of degrees of freedom that are meant to be the degrees of freedom of the zygote. And I then, a few days later, have a set of degrees of freedom that are the degrees of freedom of some kind of early stage embryo. But I want to say this is the same thing as the zygote. Well, I'm now really abusing my notation. In a really serious way by saying that these two are the same thing even though they have different state spaces. I think this is a fundamental question for any science that bases itself on the notion of a state space. How to talk about moving between different degrees of freedom and associating those things with some kind of evolution of a kind of meta-thing that we regard as a single entity. And this is one reason that at the beginning of this paper we assume that we have some single system that we divide in two. We draw a decompositional boundary. And that decompositional boundary is basically arbitrary. Or it's completely arbitrary. But when we're doing something like embryology we're looking at a decompositional boundary between the thing we're interested in, the embryo. And the rest of the universe including us that's getting bigger and bigger and bigger. And so we need a way to talk about that sort of thing. And we just, we allied that problem in this paper. We talk about a fixed decomposition that divides the entire universe into two parts that we can characterize the interaction between. But yes, underlying all of this is this much larger question of how we think about decompositional boundaries that are expanding with respect to some parameter that we call time. And of course that time is not anyone's observable time because we're trying to talk about it at the level of the universe as a whole. So we very quickly get into extremely deep questions. So two things that came up there really for me is like, okay, yeah you're thinking about the embryo. Suddenly that's like a system that's separate from the world. But even like within the embryo, like where do you draw the line between the placenta and the embryo and when? So like there's even some fuzzy blurring at the developmental scale. And also really the notion of time and time irreversibility and memory and encoding was just something I found absolutely fascinating about this paper. So you mentioned that the quantum, the free energy principle for generic quantum systems does not imply like a space-time background, like it doesn't need a space-time background to operate, but then like brought time back in in terms of memory and encoding. And I just wonder if anybody maybe wants to speak a little bit more about that or help us understand how and why you don't need space or time for its work. Well I could talk about that a little bit if no one else wants to pick it up. Let me go back to the first thing you said, W, which was about boundaries and how hard it is, for example, to draw boundaries in something like embryology. There's another deep, deep assumption that we make that the boundaries that we draw in the world are somehow ontological or objective and observer-independent. And quantum theory and many other sort of theoretical approaches remind us constantly that that's a bad assumption. That drawing, thinking of these boundaries as ontic causes all kinds of formal problems. So for example, the associativity of the state space multiplication operator becomes invalid. So you have to throw out the whole vector space form of this. So I mean real problems. So boundaries are observer-dependent in any formalism like this. So what do we mean by this universal time? It's a construct. It's a theoretical convenience. And we map it onto our observable time, which is time measured by our clock. And since we're the theorists, we can say, okay, we're going to privilege our clock. And we're going to assume that our clock is somehow universal, even though we know that that doesn't make any sense. And when we think of space-time, the term background-free comes from cosmology or quantum cosmology. And it's the idea that if we want to understand gravity, or if we want to understand space-time as an emergent phenomenon, then we need a theoretical approach that doesn't build space-time into it. But if we think of biology, how many organisms, what fraction of organisms, have anything like our experience of space? To what extent is space part of the world of organisms that we're interested in? And to what extent is space as we see it a construct of our ability to make certain kinds of measurements? And in a strict quantum-theoretic formalism, time is not an observable, but distance is. So it's an observable like any other. So in a very strict sense, quantum theory is not a mechanical theory, because it's not really about things going on in an independent space-time. Stephen, did you have a question? Yeah, I'd just like to comment on, I say thanks for referencing state spaces for systems. I think that's really helpful, because I've been wrestling with this dynamical part of systems which is really helpful to think in dynamics, but most, or a lot of people when they talk about systems, they just think about relationships between things and system dynamics. It just flows as if it's like things being passed around rather than dynamical processes. So I think it's helpful to think about state space in that way to frame systems. And I'm just curious about how we have to be careful about people who use systems in those other ways and don't have a dynamical concept of something which is nonlinear being taken into account, because that seems to be something that this approach really gives the ability to do. So I wonder what your thoughts are on that. Carl, would you like to say something to that? Yeah, I'm not sure it's going to be deeply informed. So as I'm listening, I'm trying to translate all these beautiful ideas into a classical framework that I'm much more comfortable with. The last point is, from the classical perspective, absolutely fundamental. I mean, you start with a launch valve or a rampant dynamical system that is just cast in terms of floor operators or movement operators that you can co-fex with time if you want to, but just to reinforce Chris's point. Time and space are both constructs and they're just explanations for these floor operators. So if those floor operators have a dynamics, then implicitly you've got a rate of change with respect to something which is time, but that's quite a big move. So putting things in terms of mappings, I think, is absolutely fundamental. So, Steve, if your point was when you need to think about systemic processes or systems from the point of view of mappings, usually associated with some time, I get the impression from working with Chris and Jim that there is an ordinal or a sequential aspect. So just coming back to the question, where did memory get into the game? There can still be sequences or sets of things that constitute a memory, but it doesn't have to have the attribute of time in the sense that we would understand a flow from the point of view of a differential equation. So I'd like to hear sort of Chris and Jim's comments on that. While I'm talking, though, just this importance of time as a construct, it does strike me that that takes, there's a lot of heavy lifting if we just abandon the notion of some universal or clock time in relation to what I would, from a classical perspective, read the scale-free aspect of it. I'm thinking here of the renormalization group. So I'd be interested to know if there is a renormalization group or the apparatus of the renormalization group in quantum physics, from my perspective, that notion of having a scale-free approach is exactly what is the aspiration of any, in particular the free energy principle approach would have when trying to account for the progressive increase in the scale of a system but preserving exactly the same dynamics, the same Lagrangian, the same, so again, from a classical perspective, the same functional forms for the things that are conserved. So this is a reiterate Chris's fundamental point earlier on that we're not talking about a commitment to a particular scale be it very, very small or very, very big. What we're talking about is a commitment to the same functional forms and the apparatus and the mathematics that underwrite the dynamics, if you like, and that's at any given scale. And then that calls then into question, how do you get from one scale to the other? And one thing that sort of jumps out at you when you look at the renormalization group in relation to summarizing dynamics of random differential systems is that you get this time dilation or this, you know, you move from one time scale to another time scale. So I'm coming back to the point, you know, when you talk about time, what scale, what time are you talking about? Are you talking about evolutionary time? Are you talking about the time that the fetus is enjoying in its mother's womb? Are you talking about the time scale appropriate for the fluxes of ions over a cell membrane? So the deeper question here is how does one time relate to another? The people that I know who are interested in this issue from the point of view of the perception of time or time as a construct usually fall back on information rates. They usually fall back on the rate of belief updating or the amount to which you have moved in some belief space, technically a statistical manifold from a classical point of view. You can say rate with respect to what? And of course you're back with some time, at which time, and where you end up is, well, okay, I can't talk about any absolute information rate, but I can certainly talk about the number of moves I make at one scale per number of moves I make at another scale. Even we could be doing that in our heads, you know, we have very high scales, say from the point of view of a generative model in our heads, dealing with things that unfold very slowly in relation to fast-moved updating much closer, say to our sensory organs or actuators. Now, it does make sense to actually talk about how far have I moved in terms of an information length in my belief updating at a higher level of a hierarchy or a higher scale in relation to the number of moves or how far I have moved at a later scale. I'm going to go into time, okay? Yeah. Okay. We've got somebody chatting away. Jim, do you want to mute if you have background noise, please? Okay, or you can just... It'll be muted on live stream. Continue. That little response was very rambling, because there are so many just questions that have come up over the past 10, 20 minutes. I'd like to pass that to Chris and Jim now, though, just to get their take on how comfortable they would be by taking the notions of, say, the renormalization group as applied to pathological formulations of undiagnosed systems and whether that has any homolog or any currency in a truly background-free quantum formation. No. Well, basically, I think what Carl is proposing concerning the renormalization group and van der Waals' theories is something to be investigated. But I'm inconsolent with Carl concerning the aspect of memory from the informational point of view. And may I speak about the specifics of our construction, namely the Co-Cone Diagram, which is a distributed Yeah, I'll be there. Yep. Continue, Jim. Okay, you could mute or just... Yeah. That's all good. We're going to be looking at the Co-Cone. Maybe just Chris or Carl, you can give a first pass. Like, what is this being shown and also how does it relate to active inference? Hello. Okay, welcome back. Continue with the Co-Cone. Local distraction. Yes, okay. Well, you see, at the beginning of the paper when we discussed the holographic screen, which underlies the Markov blanket, you have these preparation measurement operations which feed into the Diagram 14, of which the lower part emanating from D' is a memory write system. So as the top part is picking up information from the measurement and preparation operators, there is a memory write system being implemented from the opposite direction. One thing I would like to say about the individual classifiers that we haven't sort of fully mentioned in the paper, that an individual classifier can be the result of an algebraic operation on two other classifiers. These algebraic operations can be of the type such as concurrency or orthocurrency. For example, when you have a system of trains going through a system of stations, the classifiers in question can be combined to that extent and also by choice. So I think the scope for what is happening informationally in Diagram 14 is quite significant. Originally, choose spaces were applied to such areas as concurrency and high-dimensional automata and also we can look at the possibleistic case and there are ways of interpreting Diagram 14 in terms of the graph network too. So I know this doesn't exactly answer Carl's questions but I want to say that there is the mechanism there for the memory right. So this is, of course, influential when you see that active inference leads to minimizing the variational free energy principle. If you have any questions about those diagrams, please ask me. Perhaps Chris can elaborate on what I'm saying. Let me make a background comment here which in a sense goes to one reason that we're using this quantum theoretic formalism. If we think of everything in terms of a quantum theoretic formalism instead of a classical formalism, then the only place that there's any classical information in the system is actually on the boundary between the two interacting components. So in that original... I don't know whether it's figure one or the very original idea in the paper is that we have some closed system which we can think of as the universe as a whole or the universe of interest as a whole and it's in some quantum state and if we divided into two and we make the assumption that the joint state after the division is separable, that we've divided it in a place where the interaction of the system with itself is weak enough that we can think of the two halves independently or the two components independently, then we have this natural sense of classical information written on that boundary and that's the sense of classical information that one encounters in the holographic principle in cosmology. And it's in that sense of classical encoding that this boundary serves as a Markov blanket. But since the boundary is the only locus anywhere of classical information in this formal description, then any classical memory has to be resident on that boundary. And this actually brings to mind a picture which was not included in this paper but that I'm using in a subsequent paper with Mike Levin which really is just thinking of this boundary using a pie chart and as we pointed out in this paper another kind of classical information is the classical free energy flow that powers information processing. So this actually isn't the right picture. And so one can immediately carve off a large chunk of the boundary and say, okay, this is just thermodynamic exchange. These bits aren't useful to anybody. They're the bits that are being used, burned as power. Yeah, this is a better picture for that. And then there's a chunk of the pie chart that we can think of as perception and another chunk of the pie chart that we can think of as action. But then some other piece of the pie chart has to be allocated to memory or there's no internal time. Having an internal clock and having a memory are the same concept. If I don't have any memory at all, then I have no awareness of the passing of time and I have no internal clock. So when Jim talks about that previous Cone-Cocone diagram, a memory write operation, it's a memory write operation on the boundary. And of course the opposite partner in the interaction, my memory segment of the boundary, from my point of view, might be my partner's perception segment of the boundary or might overlap with my partner's perception of the segment of the boundary. I have no idea how my interaction partner cuts up the boundary from their point of view. So that's what this picture sort of illustrates. So this question of what is the memory also becomes observer-independent, which gets back to this question of time and the relationship of time to a local clock and the relationship of a local clock to memory. So whenever we talk about time as observed by someone, we're making reference to a local clock and a memory resource and of course an energy flow that powers that memory resource and prevents it from decaying entropically. So we keep circling back to this abstraction of the universal time that parameterizes the universal dynamics. But this formalism, like any such formalism, requires that concept as a theoretical construct. But at least it doesn't confuse it with the experienced time of any particular observer, which has to be represented in this very different and this very distinct way. Awesome. Thanks. So I think that is a perfect segue into a question that was asked by Franz Kutschling in the live chat. And also, Mike Levin has joined us. I don't know if he's able to talk. He says he's having some internet issues. I want to say hello before I ask Franz's question. If you can. You're muted. Okay, I'll ask the question and then we'll see if Mike figures out how to unmute and say hi. So Franz asks, how does the definition of a thing that maintains its identity over time scale with time scales? Or over time scale with time scales? There you go. How does the definition of a thing that maintains its identity over time scale with time scales? I still didn't get it. For example, I would call something a thing if it persisted for some time with respect to my observational time scale. So in other words, can quantum theory say something about how to set temporal observer-limited boundaries of defining a thing? There's a lot in there. So how does the definition of a thing... Can I give one first off, Lou? Yeah, yeah. I see many ways, some that are actually implicitly scale-friendly or scale-free, like the hierarchical parametric models in SPM, Statistical Parametric Modeling. That's separating, like, in a multi-scale experimental design from an experimentalist's perspective into signal and noise, including the combinatoric interactions in a classical parametric state space. And there's also the Bayesian identification of parameters, which also is, just as was brought up, scale implies choosing what the state space is. And so what statistical variance gets partitioned into the noise term or the kind of ripple term and what gets partitioned into the wave term, into the signal, into the regression slope rather than the variance term, these kinds of variance partitioning questions can be approached in a instantaneous sense, like descriptive statistics, even if it's about time series. And I think a theme that we're hearing brought up is that taking repeated measurements seriously brings in a whole host of essentially cognitive features, like, as well as information, theoretic, perspective, dependence, all of these features that are tucked under the rug or down-projected onto in a special case. If we just assume a synchronous universe or a synchronous regression, it's a totally different scenario, but considering this quantum statistical background helps us deal with the cognitive modeling and also take the consequences of action into account because action is implied in the measurement preparation, recording, experimental, active inference cycle. It was just so subtle in the paper where we said, all inference in this scheme is active inference, but the paper was about quantum and the FEP. Thanks. Could I bounce off what Daniel just said very quickly? Sure. I was also thinking about the nature of action. I thought what Daniel just mentioned there ties in with is if memory happens via there being some action, and then the action defines the memory scale. So I think that makes that starts to define a scale. So I'm curious that if there's something different between the temporality of a thing maintaining its thinness and the sort of temporality of actions by systems of things, because I think things can only act to maintain their thinness, but they don't. If they start doing things with others, they become a system. So I suppose you've got where that action still starts to come into the equation. Or am I mistaken there? Can I address that briefly? Yeah. I think the relevant fable here is the fable of the two drunks who are walking home from the bar leaning on each other. And if they're lucky, they manage to keep each other upright. If we think of an observer in an environment, and so we draw a boundary between the observer and the environment, and we think about the observer's interaction with its total environment, then we've divided the boundary up as we were talking about earlier into a pie chart where a large fraction of the bits are fuel, and so they're uninformative, and there's a perception space and an action space and a memory space and so on. Now let's think about what happens when the observer identifies a thing embedded in the environment. So the observer has now segmented or sectored the perception part of the space into a part that includes bits that represent the state of this thing and the rest of the perception space which are bits that represent the background that the thing is embedded in. So the thing's environment. And so the question becomes, what does it mean to say that that thing retains its thingness, its identity over time for that observer? Well that means that that observer is able to maintain this distinction between bits that communicate information about the thing and bits that communicate information about the thing's environment. So this becomes a cognitive question about the observer. So it's the question how does a thing maintain its identity is replaced in this way by the question, how does an observer of a thing maintain the thing's identity for himself or for itself? How does the observer maintain the ability to lock in on this thing is distinct from the environment? So let's go to an example. Think about motor babbling in an infant. So here's this infant she's lying on her back she's waving her arms and her legs around and what she's doing in this motor babbling is figuring out the relationship between the motor system and the visual system. So among other things she's figuring out that something that she sees some visual stimulus is actually part of her body. It's a hand. And that at a certain moment she begins to be able to control what this hand is doing using visual and motor feedback. So here's a living example of an observer fixing the identity of an object. And what we're trying to do here in a sense is to insist on thinking about thinghood always in this observer-specific way. So I notice that Carl just raised his hand. So let me turn it over completely to Kim to carry on with that or contest it if you'd like to. No, I don't want to contest it. I want to fully endorse that. And if you just look at the free energy principle as classically formulated it is exactly along the lines of the thingness being an attribute of the observer. So if you just start off with the notion of a Markov blanket that stipulatively defines thingness when it's defined by the conditional independence between the dynamics on the inside versus the outside but of course those dynamics are not observable once they're on the inside from the outside by definition. So you're immediately imputing some magical observer that could actually recognize that independence before you've even defined what a thing is. So I wanted to in an abstract and possibly even philosophical way reinforce this notion that thingness is an attribute of something that could possibly observe it even from the classical perspective of the free energy principle predicated on the notion of a Markov blanket that is in turn defined in terms of conditional independence or disentanglement. That example of Motabavali I think is again a lovely illustration of observation underwriting thingness and in this instance a highly anthropomorphicized thingness named itself would to develop the hypothesis or the explanation for my holographic screen or all the information on my Markov blanket that a plausible parsimonious explanation is that I am a thing I am a self I am me takes months if not years to fully actually arrive at as you're an infant slowly discovering that I am different from mom and perhaps mom is different from her background ultimately perhaps I am something like mom so the emergence of thingness I think is quintessentially observer dependent both mathematically and in terms of developmental psychology great thank you did you want to add on I was trying to lure my hand it does that by itself I think how lucky the technology how lucky the technology so maybe I will put this back to Carl maybe you were talking earlier before Jim started talking about the rate of information flow and the rate of belief updating and so I just wonder I've been curious about this actually just in terms of human information like are we really bombarded with more information now than we used to be and is this like influencing our rate of cultural change cultural evolution because as a culture like we update our beliefs more often do you think that these two things are proportional the rate of information flow and the rate of belief updating something I've been thinking about yes no I'm sure that's absolutely right and right I think at every scale of analysis so you couldn't look at it from a point of view of culture and the availability of information on social media and globalization at a cultural scale but you could also I think bring us back to motor bubbling the degree of belief updating an infant has to do is clearly going to be much greater than somebody of my age so you know time will a year in the life of an infant will not be the same as a year in my life and I think that that is a direct reflection of the degree of belief updating interestingly that question touches upon the separation of time scales which I think underwrites everything that we're talking about here so you just stand back again and go back to some of Chris's earlier observations about the the remit of the the paper in question it was really at any given but not specified scale where we could assume that that we didn't have to worry about the separation of time scales from the point of view of a cognitive neuroscientist for example what we're talking about is inference where inference in this instance as a kind of reading and writing onto this holographic screen or if you're one of the classical perspective the updating of active and sensory states on the Markov blanket but under the assumption that other things that change at a slower time scale are not changing and those other things might be for example at a slower time scale they might be learning so you've got this notion that the formalism at hand the reading and writing at hand from the Coctham information theoretic perspective could be a very good metaphor for inference over a short time period where things are not changing very quickly from the point of view of the things to be learned so within a few seconds or minutes of the life of a zygote acknowledging that you could apply the same apparatus and the same mechanics at a slower time scale that would account for the transition from a zygote to a neonate as it learns so the difference between inference and learning I think is a nice common sensible example of this separation of time scales so when you ask about cultural information available I think you're probably talking about a scale of learning absolutely when we learn we update our beliefs there it will be from a classical perspective a statistical manifold that holds those beliefs and you're literally moving over that and with respect to another scale then the degree of belief update and the degree of asking questions the degree of rewriting will certainly be a function of the amount of reading and writing that's going on and I can imagine somebody who has access to Wikipedia or news or social media is doing a lot more rewriting than I will say in the 1960s you're muted but to me blue yes to you Carl you made a fascinating point about connecting the motor babbling and taking it into the informational and the belief updating domain into the adult where we're kind of motor babbling on social media in a way that has different dynamics than before and people often give the example of like well a year feels longer for a 7 year old because it's the 7th of their life but then for the 80 year old it's the 7th of their life so that piece of the pie is smaller versus bigger and that is implicitly the fallacy of the observer independent time as if there was a linear chart that was being linearly divided by some static observer whereas another way to look at that or an alternative model is like the rate of development or experiences or change or novelty or continue to explore and don't want to all mute James he's muted on the live stream so that's all there is that rate of belief updating could set that internal kairos, that biological time like that biological information geometry and so it connects the cognitive sciences and also the statistical perspectives that have been discussed to this time question as also brought up earlier does so much work in so-called dynamical modeling and it just makes me think about how there was a movement from the observer independent synchronous into the relativistic which prepared and almost like prepared the partition to exist so that all these other domains like the holographic principle could come into play so my question for any of the authors is where does the QRF come into play and how does this relate to discussions of pluralism and polycentrality awesome question Jim did you want to answer? I wonder if you can bring up a diagram there in the paper let's see yes it's diagram 34 there you have a configuration that relates the QRF to the focal cone diagrams and we say here that let's see well here we focus on co-deployable observers and this diagram commutes it's interesting of course when the diagram doesn't commute and then you have a kind of intrinsic context so I think what this is saying is that there are certain elements of information represented by the logic of the distributed system in question that are kind of ambiguous or rather they're manifest in classifiers and infomorphisms that aren't defined or the co-limit the bold C at the top there does not exist so this is a very useful criteria I think to distinguish between sort of empirical models which the best part are based on co-deployable variables and what happens when a quantum contextuality occurs that is diagram 34 doesn't commute so there's some sort of discrepancy there in the mechanism of the quantum reference frame relative to the dynamics of the Markov blanket or the holographic screen underlying it so there is some kind of distortion in meaning that results in this sort of intrinsic contextuality I don't know if that's a satisfactory answer at all but when this breaks down it sounds good Chris I'm sorry there's another local distress Chris did you want to add on to that or make another comment I wanted to go back a little bit to the basics of what a quantum reference frame is because this is a concept that is both very familiar and I think a little bit unfamiliar and the term was only coined back in the 1990s when the idea of a quantum measurement was associated with the idea of the units in which the measurement is made and what are the requirements for a measurement being meaningful and we're going to get right back to the same issues we've been talking about with respect to time and thinness but a reference frame is just a way of giving meaning to an outcome and that's what allows outcomes to be compared so if I have an apparatus like a meter stick I can go around and measure lengths and as long as I believe that my meter stick maintains its identity as long as I believe that my meter stick remains the same thing then I can compare the lengths that I've measured but if I don't believe my meter stick remains the same thing then these length comparisons are meaningless so the notion of a reference frame is intimately tied up with this notion of meaning which in this free energy principle context is really a notion of actionability can I use what something means to guide my actions in my decision making so this notion of a QRF becomes a notion of giving an outcome value that I obtain and write in memory, comparability with other outcome values that I've maintained and write in memory and we're very used to, as human beings we have this incredibly complicated sensory and inferential apparatus that involves many many degrees of freedom and many many values that are distinguishable so I think if we want to think about something like the role of reference frames and perception it's better to think in terms of something like E. coli and so think of a much simpler kind of system that can measure fewer things E. coli can measure concentrations of salt and sugars and various kinds and those concentrations are meaningful to it those concentrations allow it to make decisions because it's internal biochemistry assigns them an actionable meaning so chemotaxis and E. coli works because the level of phosphorylation of a particular chemical remains reasonably constant that's the reference frame and concentration levels can be compared to each other as long as that phosphorylation state remains reasonably constant that reference frame is meaningful and so chemotaxis makes sense because E. coli has if you will a slow clock the phosphorylation state of QI that assigns that confers meanings upon its outcomes so now we can spool back to what Jim was saying about co-deployable reference frames reference frames that I can use at the same time that I can obtain information from simultaneously have to commute this is what Heisenberg's uncertainty principle is about and if I can't use them at the same time it's because they don't commute if I use them in different orders I get different answers and what that means is that by switching reference frames I'm effectively changing the calibration I'm effectively changing the meaning that I've assigned and I think a very beautiful place to see this is the paper, a paper that we referenced here from Cervantes and Zaffer and Zafferoff called the Snow Queen experiment the paper is called Snow Queen is Evil and Beautiful came out a few years ago and basically this paper is demonstrating quantum contextuality and human decision making and it's doing it by exploiting differences context dependent differences in the meanings of words so it's perfectly pointing out this relationship between meaning and co-deployability and hence context dependence in human cognition so that's a paper I very much recommend reading thank you so much Carl did you want to comment on that also? I'm aware that Stephen also has a question but just to follow up on that beautiful deconstruction this is more of a question but I'm going to state it as an observation this notion of the QRF as being fundamentally getting measurements right or metrics right would be understood from the point of view of inference technically as getting an internally consistent metric correct and the metric we're talking about is a Fisher information metric so if you look at all reading and writing or action and sensation from a classical perspective as just trying to minimize some divergence or some prediction error what you are saying is that you are trying to find an internally consistent metric in an information space that makes sense of your active sampling in the world so it just sounded to me as if what Chris was saying was exactly getting the measuring stick at least behaving consistently you'll never know whether it's changing its length or not but at least if it behaves consistently in relation to what you can see of the world the way that you can query that world actively then that's good enough and that of course simply stating that the name of the game in active inference is just to minimize your prediction error where you are now reading prediction error as a KL divergence and of course the information length or the amount, the measurement or movement the metric is just a path integral of infinitesimal changes in KL divergence so it's seen from a mathematical perspective the perspective of information geometries there's a beautiful link here between aligning your QRS and getting your generative models right to make the most sense in the sense of minimizing or making as internally consistent as possible all your information metrics Awesome, thank you Before we go to Steven's question we have a question or two from Andrew Aguirre in the live chat he asks does the existence of a holographic screen slash markup blanket automatically yield a statistical manifold for the belief state of the agent and is there a canonical renormalization procedure for a quantum mechanical system? If there is one how would the core screening translate into the statistical manifold? I can take one and whack at that which is if I'm core screening something if we think of this in terms of QRS then I'm deploying a different reference frame with lower resolution so in a sense what's being renormalized is my representation of the units that I use to distinguish states of the world so if I have a meter stick that's made of stainless steel that has millimeters indicated on it then I can go around and make measurements in millimeters but if I course grain to a meter stick that only has half centimeters indicated on it then I've lost a factor of five in resolution I've course grained by a factor of five and so I've in effect rounded off all my outcome values and now my function of comparing outcome values is differently defined because it has to work at this lower resolution and if I want to compare my previous values to my later values I have to incorporate this explicit round off function into my comparison of values because I've changed my QRF so course graining which of course couples to scale changing from the very beginning of our conversation can be thought of as a transition from a very QRF with some grain size to a QRF for a similar degree of freedom different grain size and that has ripple effects through everything including the free energy cost of memory right if I'm writing two bit numbers into memory then that's a lot less expensive than writing five bit numbers into memory so this raises another aspect of this question of renormalization which is how does it influence how can I use it to solve the trade off problems that arise if I'm an organism with a limited fuel supply which of course I am going about the world trying to gather preferentially the gather the information that's going to be most useful to me at the lowest possible cost so again we get back to this key issue in all of active inference of balancing the risk of exploration against the reward of new information and coarse-graining is intimate to that and it involves the choice I make of the tools that I'm going to use to make measurements and how that choice impacts my need to allocate computational resources and energy resources to what I'm doing to the complex task of getting new information while at the same time surviving awesome thanks Daniel Carl do you want to speak to that first? perhaps Steve should ask but then I can come back to what you just said yeah I'll make a note and then Carl it would be great to have your response then Steve into your question so to transpose it into statistical parametric mapping the true classics there's a mapping of measured voxels which are being defined by the resolution of the EEG or the fMRI machine that's the holograph that's like what the experimental apparatus is providing is the measurements those are the O in our partially observable Markov decision process it's the predicted value in the regression and that space you know if it's 128 channels in the EEG and they're getting this sampling right is the state space that's being discussed and that's being transposed or mapped or connected to a different statistical space with different features and rather than the voxels there's like the mapping onto resels so Carl that's why I'd appreciate if this is true or not but those are the resolution elements and that it can be like a coarse graining and in that stretch meter stick and different space there's a lot more statistical inference options available that are just not accessible at all from the voxel coarse graining which is noisier and there's a signal enrichment in the coarse graining to the resels that's appropriate and then that gets juxtaposed with the contrasts which is the experimental design of the scientist like you couldn't differentiate A versus B category if you only tested A so the experimental design and that's where the statistics actually comes into play and then just as Chris said there's the organism with limited food supply and then there's like the scientist with limited time attention computer storage, processor time, funding maybe seven billion people like just there's plausibility and so that's where this operational co-deployment connects with experimental design and so many of the patterns and analogies exist in the purely classical domain of the parametric statistics like SPM in the classical physics and the flow and the information flow area that's coming into play and also these more recent in quantum areas so Carl would love to hear your thoughts on that. Yes of course I'm mindful that Stephen still has a question but perhaps I'll speak to that and then Stephen can ask a question but I think yes these questions and Daniel your responses are touching and Chris is touching on something really fundamental so just to answer the question from the point of view of Markov Plancki, yes it does automatically yield a statistical manifold or at least it looks as if there is a statistical manifold if you can observe the behavior of the thing that you are observing. The second thing question which we've been focusing on and the course screening I think that's a really fundamental thing. I would argue that not to conflate too much the notion of renormalization which would be I think more a question of how you get from one scale to another scale with the notion of course screening which could be read to simply as the degree of quantization or choosing the right resolution element, the right course screening of a given QRF at any given scale you could argue that you could contextualize it but I think there's a more fundamental argument here and that's the argument that there will be from a statistical perspective an optimal course screening for any given holographic screen or exchange over a Markov Plancki and that's the one that statistically maximizes the marginal likelihood which you could if you like relate to this sort of getting the measurements as internally consistent as possible maximizing the predictability of the next measurement you make that goes hand in hand with a maximization of the marginal likelihood or a minimization of the variational free energy that is important to remember because the marginal likelihood can always be expressed as accuracy minus complexity where complexity scores the degrees of freedom you're burning up if you remember about the bits on the holographic screen you need energy from a classical perspective represent what that translates to in a very simple way in terms of probability distributions and functionals is that in maximizing the marginal likelihood or the evidence for your aligned or model of the world you are to maintain a certain degree of accuracy trying to minimize the complexity so you're using as few degrees of freedom your course screening to the extent you can get away with while maintaining an accurate account of the sensory states or the perceptual space on your holographic screen and now from the point of view of active inference just think about what would that accuracy and complexity look like in terms of the consequences of action you get exactly what Chris was talking about which was the expected complexity being the risk and I mean risk in the sense of an economist or basin decision theorist the accuracy now becomes a sort of expected utility or the expected log evidence, the expected likelihood that you would get if you committed to this kind of course screening so this course screening is absolutely fundamental it's really interesting to go back to all the sources the course screening starts to make a difference and I'm thinking here coming back to Daniel's notion of resolution elements you find this in sort of vena filtering there is an optimal level of smoothing or blurring for any given kind of data about which you want to make an inference or you want to assimilate you see this in formulations of universal computation that themselves are based upon induction that get you right back to complexity and compressibility so you see it in the drive that underwrites all formulations of universal computation to maximize compression by minimizing the complexity of your message passing getting as efficient and as simple as you can this best way to make your account as simple as possible is to course grade it and I think you also see it in the different levels of cognition in psychology and in neurophysiology in the sense that the higher level more efficient ways in which we align our QRFs or optimize our generative models always lead you to a very course grade quantization of the world so for example I represent I am sure in my brain on some neural populations the fact that I am in my study I do not represent that in terms of very fine grained XYZ coordinates down to the millimeter I represent it no I am in my study I carve the world at its joints and those joints basically demarcate the kind of course grading I bring to the table there are all sorts of arguments you can pursue here to vindicate this is the very existence of receptive fields tiling a sensorium in a way that is covered by discrete quantized receptive fields tells you immediately that the brain represents stuff on the inside if you can look at it through brain imaging is in this quantized fashion this course grained fashion and that course graindest and the size and the deployment and the organization of those receptive fields has been optimized at many different temporal scales so I just give you a few takes on course grading as something absolutely fundamental to the structure of a generative model or to the way that we align our QRFs to make sense of the stuff on the holographic screen at least when we are when we are reading from that and sense making in that sense Great, thank you, Stephen Yeah thanks, this is really cool I was going to come back a little bit to now and the past and the idea that you mentioned about the way that we interface or the assumptions of the interface then changes how we resolve the thing and the background sort of situation situatedness of that thing and so that could then also and this is the question in a way is how can that also because if the way that that thing and the assumptions are made sets up like the meaningful influence the action policy selection and the way things are thought about how can that itself change the way that brain or the way that organisms look at what things are for instance in the modern world we do see everything as things we carve things up, we see things as separate like an individual is separate and that's the sort of indigenous approaches time for them was cyclical or followed the river going down a hill so there was a complexity focus in a way, it was almost like the mounting car problem of finding the landscape of meaning rather than defining the thing and what it is and carving it is how much is the way that the quantum reference field translated into thinking about what that thing is and how that compares to say our western cognitive frames which kind of we measure everything because we can because basically we can control our environment so heavily we can make things very accurate whereas in the past it might have been the Kalahari Bushman or the people in the middle of the Amazon that somehow go with the landscape of complexity so that the way that is understood is going to be different so I'm just curious about how you think about that ways of knowing Great question, I wonder also if that ties into the individuality of temporal scales, like you mentioned the cyclical time and back to Franz's question maybe earlier about how a thing over time that stays a thing over time scales with the scale of the time or with the cyclical nature of the time or it reminds me back of when we had Chris and Shana here talking about like flattening time maybe that was just Shana's talk but yeah anyway I don't know do you guys anybody have any temporal comments? Just that there's an analogy to the synchronization of clocks which is a huge part of computer science and of quantum and other areas that question of synchronizing clocks seems to have some relationship to cognitive model not naive congruence in terms of synchronization but in terms of like time zone coordination so if we're in the same time zone then there's a comparability but for cognitive states and so I think that will be something awesome to explore in next week's session and of course in the long run about what that kind of semantic and rhetorical and narrative alignment that isn't congruence but reflects a broader frame which actually the differences within the frame are the search through that field Thanks and in the interest of temporal comments I think maybe does everyone just want to go around and give a final thought and maybe an idea of what they would like to see in next week's discussion and we will start with Steven Yeah thanks lots of interesting things I suppose I'm curious about the way that collapsing of the quantum reference frame might scale differently the assumptions that lead to the collapse could inform different scales of interactions and mixing and blending of behavior and digesting a lot of stuff that's come up so yeah thank you for the opportunity to talk to you Awesome Daniel anything you want to see happen next week We always think of the dot zero it's kind of like dropping into the skate park or the bathtub and so that was like a microcosm of all the threads that get brought together and then in dot one I feel like we opened up so many areas of possible connection and so I'll just look forward to the second repeated measurement of this semantic or regime of attention and see how the continued interaction and engagement updates our generative model past the dot two even Carl or Chris do you want to give a thought about what you'd like to unpack in the next question Should I go first? Well Chris thinks of something very clever to say So it was pulling on from Daniel to a certain extent Steven's question as well I do hear about the principle of unitarity entanglement and its relationship to this notion of synchronization and the mutual alignment of QRFs between things that are coupled so coming to regimes of attention I think there could be a very transparent link there in the sense that contextualizing your QRF I think for a cognitive neuroscientist would be getting your attentional set right and possibly implicit course grading that comes along with a particular attentional set taking that notion to tending to others and synchronizing and communicating and living in a universe that is also composed of things like me and bring some fundamental questions about specifically generalized synchronization quantum theoretically entanglement and enabling that entanglement in the right kind of way via aligning your QRFs and getting the course grading right in sympathy with your partner So I'd like to hear about the principle of unitarity entanglement and how it relates to classical notions of communication mutual inference and generalized synchrony Well thanks for stealing what I was going to say I was also going to say how there's one like the final concluding part of the paper talks about how the FEP is asymptotically the principle of unitarity and I really would like to dive into that more and maybe any potential implications of that and now Chris gets the final and cleverest word Okay this won't be terribly clever What I was going to do was to recommend a lovely paper by Alexi Grinbaum that we refer to in this paper and the closing sentence of that paper is that physics is really about languages and I think this ties in nicely to some of the issues that Steven was raising in his last question which I found very interesting about how do conceptual schemes evolve at the cultural level and in hearing about the whole issue of reference frame synchronization or reference frame alignment and how that relates to entanglement this really is a question about language and that paper by Alexi Grinbaum I think makes this point very, very beautifully so that would be a nice thing to look at at some point let's see this is that's probably a newer paper by him referring to is titled something like how device independent methods changed the nature of physical theory or something like that was published in studies and history and philosophy of science 2017 or something like that Anyway thank you very much for this Yeah this was very interesting and engaging and really appreciated everyone's participation and blue's facilitation so we'll look forward to seeing anyone just drop in or out when works for you next week at the same time in number 40.2 so till the next measurement Thank you take care