 All right. Hello and welcome, everyone. This is the Active Inference Livestream. I'm Daniel Friedman, and this is the Active Inference Livestream 007.0. Welcome to TEAMCOM. TEAMCOM is an experiment in online team communication, learning, and practice related to active inference. You can find us on Twitter at inferenceactive. You can find us at active inference at gmail.com. via our public keybase team or on YouTube. This is a recorded and archived livestream. Please provide us feedback so that we can improve our work. All backgrounds and perspectives are welcome here. Video etiquette for livestreams is that you want to mute. If there's any sound in your background, raise your hand if you want to speak so you can be added to the stack, use respectful speech, etc. Today in Act Impstream 7.0, I'm going to try to set the context for 7.1 and 7.2, which is going to be a discussion of the paper Variational Ecology and the Physics of Sentient Systems by Bramsted et al. 2019 in Physics of Life Reviews. This video, as the other .0s are, is just an introduction to some context. It's not a review or a final word. The goal is to provide context for the ideas, philosophical threads, and vocabulary of that Bramsted et al. paper. Go through some of their points as well. And the punchline is that ecosystems are multi-scale systems with implications for how we study ecological organization. And the sections of 7.0 will be as follows. First, we'll go through the keywords and talk about some of the publication context. Go through the goals of the paper, the abstract of the paper, and a roadmap of the sections. Go through some key quotations, and then talk a little bit about ergodic systems, some critiques of the free energy principle, the simulations and historicity and how that influences how we think about ecosystems, and also the cog ergo sum. Then in 7.1 and 7.2, we're going to discuss this paper with whoever wants to participate, so you can also save and submit your questions or get in touch with us if you want to participate. Also, and thanks for the suggestion to feedback, was to how to relate this to previous discussions. So if you want previous discussions to listen to that relate to this conversation, 1 would be active in stream 2, is the free energy principle a formal theory of semantics because this discussion and paper was asking, can we use the FEP to study meaningful systems? So both the ones that matter for us, the meaningful systems for us, but then also the meaning making, the auto poetic systems. Another discussion was 5.0 and the 5.1 and 2 discussions, which are the context for this question about internalism and externalism, which is a question about how and where action and perception occur. Are they internal to an organism or do they incur in this space that's mediated between the external environment and the internal environment? And then also this most recent 6.0 and following discussions where we talked about tail of two densities, active inferences and active inferences, because there we talked about how inactive or ecological approaches are broadly linked with computational approaches, like does one have precedence over the other? All right, so there's a lot going on in this variational ecology paper. So we're going to start here with the keywords. So there'll be pages for each of these, but here's the keywords that were on the paper, niche construction. So niche construction is when some system of interest, like an organism over developmental or ecological or evolutionary time scales, modifies its environments. It's pretty broad and we'll talk more about it in a bit. Variational ecology. They're using variational in a sense here to mean just broadly multi-scale and then akin to the processes of variational Bayesian methods, which are kind of like general multi-scale methods. So they're thinking about ecology from a multi-scale perspective with a sort of flexible computational aspect, as we've seen from those previous discussions. Evolutionary systems theory is basically just evolutionary biology. Probably some people prefer evolutionary biology or evolutionary systems theory, not exactly sure they're pretty equivalent. Physics of the mind, we'll have a slide to discuss because there's a few different senses of what physics of the mind entails. Free energy principle, of course, one of the keywords here because that's sort of the big framework or the paradigm that everything is being worked on or responded to under. And then variational neuro-ethology, which we'll talk about more soon, but variational again here is just meaning multi-scale, ethology, no, it's not Ethereum, it's actually behavior, and neuro, just meaning the neurophysiological component. So this is kind of like neurobehavior and thinking about the cause and effect between the neurophysiological parameters, whether they're neural or hormonal, and the behavioral outcomes and the collective behavioral outcomes, multi-scale. All right, so let's turn to ecology first. Here's a quote from the paper. They wrote, In summary, the niche transcribes regularities that pertain to group behavior. They are transcribing the physical layout of the niche through the active and perceptual dynamics of the agent, which coincides with perceptual and active dynamics of the niche. The same dynamics entail the landscape of action possibilities, which are affordances, which maintain the structural integrity of the agent-knees system. Knees dynamics resolve the free energy gradients that are induced by the physical structures of organisms in their niche and their history of dense interaction and dynamic coupling. In this sense, the robustness of patterns of shared intentionality and an action of shared meaning is inherited from the robustness of niche and vice versa. So to kind of ground this in some other more non-FEP ecology, here's a 2014 paper by Deborah Gordon. And here's two figures that illustrate this point that I'll make, which is that there's going to be sort of this hand-in-glove relationship between the dynamics of the ecosystem and the resulting collective behavioral and, quote, individual behavioral algorithms that exist. And first off, the variational approach or just the multiscale approach means that it's kind of collective behavior all the way down. Like, inside of that ant on the bottom right, there's also a collective of neurons or, you know, it's just going to be that type of interactions of different scales all the way down. That's sort of the variational realization is that it's all going to have to be at least coarse-grained at the high and the low end, if not tried to be modeled explicitly. But we can at least draw a border of analysis around an ant and track it as something that we can identify as a functional component of the system. And so in the two figures, there's two examples of axes of ecosystem variability. So on the right side, figure one, it's patchiness in space and time. So in A, where the resource is distributed really evenly across space, then there wouldn't be any type of mass recruitment pheromone because there'd be no reason to bring an ant to where you just got that one little blip. Whereas under ecological regime B, it would make sense to have something like a recruitment regime because it could be an exploitable resource that might be patchy in space and time. And you might see ants that could switch different strategies. However, one ant species may specialize on one or the other. In figure two, it's an example of a high versus low operating cost environment. So when operating costs are high, things are tending to be off until they're stimulated on, like a fever because operating costs of a fever are high. Whereas when operating costs are low, it's an example of the maintenance of leaf cutter ants under certain conditions in the rainforest to maintain territory versus the bio energetics of moving, let's just say under one example that is often brought up here, until there's a disruption, there's a continuation of function. So this is sort of a similar realization to what is described in the paper, that there's this statistical co-variation between the collective behavioral algorithms and the niche regularity. Here is another paper by Professor Gordon in 2011 called the fusion of behavioral ecology and ecology. So on the left are just a bunch of animals, a bunch of things that people have studied. And then ethology again, it's just kind of a funny word that is used by some behavioral biologists, but not all. But this is kind of an example of the different range of things that people have worked on in that area specifically. And one cool quote from this paper was, until someone discovers a fossil record for behavior, we can in practice only investigate what are the ecological consequences of a trait now. That is the evolutionary ecology of behavior. So we can sometimes see, for example, this, you know, an amber that traps one insect, capturing another one, we can say, okay, this pickup move may have happened one time or it may have been frequent. But all we have from behavioral observational perspective are actually the live realizations of today. And so the implications are that because the local rules of engagement in ecology can change, it appears that selection shifts in space and time. So we know that selection pressures can go in one direction or have a certain intensity on one trait and then change in a different year or in different time. This is why behavioral ecologists who studies the ecology of a trait in one time and place will find the outcomes different than another. The main project of evolutionary ecology is to understand when and why that happens. The effects of local conditions on species interactions create different ecological pressures in different places. This produces a spatial mosaic of diversity. And then another quote there was, context dependence requires us to redefine the project of relating behavior and fitness. It is not realistic to hope to find any reason why any trait would always be adaptive. Any trait behavioral or otherwise is deeply linked to others genetically, developmentally and functionally. And if so, if no trait has ecological impact independently of all others. Moreover, even if we can ascertain what variation in certain behavior is associated with variation and reproductive success, changing conditions will still shift around that relation. So that's non-stable or non-stationary fitness landscapes. Adaptive landscapes are full of small bumps and the behavior itself modifies the landscape. If we watch long and carefully enough, we will see different consequences for reproductive success of the same behavior. The project then is to learn how conditions determine the ways that behavior contributes to variation in reproductive success. So behavioral ecology already there's the realization that it's important to think about this interplay between features of the organism as a ecological unit, a neuro behavioral unit, but also at the ecological and population scale and evolutionary scale. A little bit more specific about niche construction, drawing a figure from this paper by Kendall et al. Niche construction is a pretty broad term and a lot of outcomes and processes can be placed under this framework. It has to do with the way that organisms modify their niche and this is almost always going to be the case when you have conspecifics in individuals of the same species because almost by definition they're going to be taking up an overlapping, you know, set of the canopy elevation in a rainforest or foraging preference. And so this diagram shows how you can have natural selection operating to shape based on instantaneous relative fitness, the distribution of the strategies or genotypes and phenotypes on the right, this great line coming down here, that's genetic inheritance, but then also you have the outcomes of collective behavior, which are influenced by the changes in the genomic and epigenomic and niche and developmental factors and cultural factors, then influencing the way that natural selection shapes that. And that can be through semantic information or through physical resources like your social insect colony. So one example here of a cool niche construction, this is a clearing in a rainforest and the clearing comes from these ants that trim, that live within a specific plant and trim the ground around it. So here's the paper, The Devil to Pay, a cost of mutualism with a mere Militista shumani ants in Devil's Gardens in increased herbivory on D'Roya and Heresuta trees by Frederick E. Gordon. So like this paper is saying there's ecological trade-offs. So it's representing a different strategy and there's obviously a reason why these Devil's Gardens don't just proliferate and sweep across the whole area. They clearly have limiting factors. So it makes sense to study the bigger context and understand what these limiting factors are and where they break down rather than just say, well, it's just simply a good strategy. So it's always going to be about kind of a bigger level of analysis as well. All right. Kind of broadening from ecology to ecological psychology. This is related to some often human-centric obviously as psychology might tend to be modeling of systems around people. And we were all children and so there's a developmental aspect. There's also these immediate to indirect aspects of our system. So there's a multi-scale and a multi-time aspect to ecological systems theory. And also on the more direct psychological side, there's this three-way diagram that comes from a 1976 paper. And it's sort of like a cycle. You have the starting on the bottom left. The schema is what directs action and what action accomplishes exploration. What exploration does is it samples. What does it sample? It samples perceps like ocular motor sampling and stimuli. What does that do? It modifies your schema. What would be identified in the free energy as the internal model? And ecological psychology also is where we get the development of ideas like affordances and the fields of affordances, the skilled intentionality framework, what the organism or system can do. And that can reflect technology as well. So in our constructed niche, that can be technological embedding. And this image is nice because it shows the clarity of the phone, the hyper-reality of the phone in relationship to the blurriness of the embodied. And so it captures that what's at hand, what we can act on, what we're using for information, that can overwhelm our field of affordances. One term that I definitely had to look up and think about a few ways that it applied was this physics of the mind. I was just like, okay, what is that going to mean? And I think that there's a few different ways. I'm happy if someone has another way to add or if there's a more specific way to state it. But one sense of physics of the mind is that there's a search for lawful rules and simple equations for the mind to make neurosense in some sense, more like some people's conception of physics. So that there'd be a laws of motion or an invariance principle or symmetry principle or a relationship with some other physics equation for the mind. Another branch of physics for the mind is that whether or not it is what the brain really does that this branch of engineering oriented analysis of neuro systems would facilitate the development of systems engineering and complex systems and control systems. So kind of like applied physics being engineering, that's sort of where these micro sensors and brain computing interfaces are at is at the applied physics level. And then another sense of physics of the mind is that there might be some understanding of the physical basis slash substrate of the mind, which is kind of the hard question of consciousness. And it also relates to questions like soft matter physics and crystals, quantum, microtubules, what kind of quote computation is being done in the brain? What is the brain doing? These are all kinds of questions that is, you know, is it an antenna or a processor? Is it which metaphors and which aspects of them are instructive in that regard? And the big question that all of these senses of physics the mind are getting at is because the mind is explicitly a metaphysical thing as opposed to the brain, is can metaphysics join with multi-scale systems, neuroscience, physics, molecular genetics, evolutionary ecology, et cetera? Then they mentioned, you know, in the title, that it's a variational ecology, physics of sentient systems. So what are sentient systems? Because initially I think I had the wrong reading here. So VNE and the VANC, which we're going to get to in a second, are about enabling the application of the FEP to phenomena within and beyond the brain. These approaches hold the promise of extending the variational free energy approach to the dynamics of sentient systems, i.e. systems with sensory states across spatial and temporal scales. So this is not a paper or a discussion on consciousness or awareness. This is going to be about emergence in multi-scale biosystems, which is pretty compatibilist with either interpretation, but this sentient doesn't mean awareness, it's not about tearing tests, it's about sense systems. So this is about being able to model simpler sensing systems, you know, thermometers and that type of control system in a very simple way, but also more nuanced systems like other organisms in a more rich way. All right. So how do we get here? What is the literature context for this variational ecology paper? The FEP has recently been leveraged to furnish a fully generalizable meta-theory for adaptive behavior in sentient systems across spatial and temporal scales called variational neuro-ethology. So let's talk about what is VNE, because this paper is going to reference it a lot, what is VNE and what scales or types of phenomena does it model. So we're going to get to VE, variational ecology, via VNE, variational neuro-ethology, by contextualizing within a critique that leads to this paper. So in this Life Review of 2018, there's a paper by Rampsteaderall answering Schrodinger's question, which is what is life, 1944, a free life formulation. Now that paper led to a bunch of responses, which led to the clarification on some of these responses and sort of the drawing out of this variational neuro-ethological approach. And one quote that arises from the paper, the Schrodinger's paper, was, given the success of this explanatory framework in biology, we suggest that Tim Bergen's levels of inquiry might be apt to elucidate structural laws that supplement the general principles provided by the FEP. The FEP describes a general biological modeling imperative while Tim Bergen has offered a distinctive but complementary framework that allows us to develop substantive explanations for the phenotypic traits and behaviors of any given species or organism. And so here is the 1963 paper by Tim Bergen and then a 50 years update by Bateson and Laud. And we talked about this in one of the previous discussions, but this is a two-by-two grid that represents Tim Bergen's four quote, four quote, wise. It's kind of like Aristotle's four wise. It's about having multiple different kinds of explanation that are at different types and different scales of analysis. So the historical how is how did this organism get here is development and how did it get here in the very short term, the contemporary how is like the ultimate mechanism. How did my arm get here because it moved from this other place? Ultimately, a little bit more historically, I got here because it was an arm butt. And then the contemporary why is the function? Like what does the thing do right now? Just at this moment, like why is there a mouse on my desk so I can use it? And then historically, why is there a mouse on my desk? Why is it as opposed to not using a desk or not using a mouse because of evolutionary processes? So that's a little bit of a broader level question. So that kind of compatibilism amongst answers is going to be complemented by a way specifically connect multiple levels through potentially even more formal or modeled or testable ways of thinking about the relationship between these four categories. Thus constituting a complete answer at least within this four-fold framework that's been very helpful. So that is where variational ecology comes in. And we're going to talk about it in the context of building on VNE and what that is too. So the paper is Variational Ecology and the Physics of Sentient Systems in Physics of Life Review 2019. Rammstead, Constant, Badcock, and Friston. So the goal of the paper and again we're going to return back to these things is specifically we review a framework for modeling complex adaptive systems for multi-scale free energy-bounding organism niche dynamics, thereby integrating the modeling strategies and heuristics of variational neuropathology, VNE, with a broader perspective on the ecological nestedness of biotic systems. We extend VNE beyond the action perception loops of individual organisms, i.e. active inference, by appealing to the variational approach to niche construction, VANC, to explain the dynamics of coupled systems constituted by organisms in their ecological niche. So the big question is, how can we make a multi-scale ecology framework that builds upon eco-evo-devo and also adds insights from physics and mathematics and complexity? So first part of the abstract. This paper addresses the challenges faced by multi-scale formulations of the variational or free energy approach to dynamics that obtain for large-scale ensembles. We review a framework for modeling complex adaptive control systems for multi-scale free energy bounding organism niche dynamics, thereby integrating the modeling strategies and heuristics of variational neuropathology with a broader perspective on the ecological nestedness of biotic systems. So as stated in the second part here, it's going to be a framework which is theory plus tools that's going to be about modeling complex adaptive systems as control systems. What are they going to be controlling? Something about the free energy boundedness of their own interactions with their niche. So rather than just thinking about how the physiological or the neuro aspects are related to constraints on behavior, it's now also about behavior looking up. We extend the multi-scale variational formulation beyond the action perception loop of individual organisms by appealing to the variational approach to niche construction to explain the dynamics of coupled systems constituted by organisms and their ecological niche. We suggest that the statistical robustness of living systems is inherited in part from their eco-niche as niches help coordinate dynamical patterns across larger spatial temporal scales. We call this approach variational neuropathology. We argue that when applied to cultural animals such as humans variational ecology enables us to formulate not just a physics of individual minds but also a physics of interacting minds across spatial temporal scales a physics of sentient systems that range from cells to society. So briefly here's the roadmap. Section zero is the introduction. They first talk about the variational or free energy formulation and specifically about how active inference models and generative models are related to it. We're not going to go through the figures in this discussion though. The second part of the paper is about variational neuropathology which is about how ensembles of Markov blankets underlying the organismal level facilitate the organism niche environment that was sort of the neuropathology level of analysis. Then they take that sort of organism upwards view of niche construction with variational approach to niche construction of inferences and reach the variational ecology which is really just connecting these two at the organism and just making it a multi-scale ecology and then subsuming the behavioral component within it. Alright. So to get to variational neuropathology especially if this is new learning about active inference relates to other areas in neurobiology. Some of the comments about variational neuropathology from other authors in response to the articles and just points and comments that are really sometimes apparent other times not in the literature but areas for improvement in theory. One thing that's cool about the active inference area is it really draws people back to these fundamental questions that really were long term about top down or bottom up forces for example in ecology or behavior different the role of historicity ergoticity. These are long time debates but now coming into this new area where we can at least address them head on. So let's think about that perspective and I'm trying to think about how coming from a regular behavioral sciences perspective this might play. So could this reflect a fundamental distinction between free energy principle as an explanatory principle for the clearly bounded ergotic biological systems in the greater complementary forces at play that constrain complex adaptive systems in general including groups of organisms in their ecological niche. Should we restrict the FEP to level the organism and explore how the model connects meaningfully to other key concepts in evolutionary systems theory about the agent needs relation. So that's restating the question whether the FEP for some reason would stop applying at the level of the organism and then it would break down at higher levels. The aim of this paper is that they're going to address the challenges these comments that people brought up about the VNE by addressing how it sits in broader level questions. And so there are things that the field of behavioral sciences has also been thinking about. For example you can watch one bird's behavior it's what you're going to be tracking but then it's going to be related to everyone's behavior so at some point it's going to be like studying collective behavior. So what they're going to do is add the variational neuro-ethology which is the multi-scale plus the neuro plus the structure plus this niche instruction just to additively create this variational ecology new acronym new PAPE why not. By putting together ensemble dynamics or collective behavior eco evo divo and niche construction complex systems free energy principle multi-scale systems and active inference and nestedness and embeddedness and then kind of just think about how it could apply to systems that persist far from equilibrium within this flow state valley of things that are small and fast on the bottom left to big and broad geospatially on the top right. But we won't go through the figures in this talk. All right to date VNE has this is before this paper to date VNE has only provided a principled method of analyzing nested and mutually constraining biological systems and their complex adaptive dynamics across spatial temporal scales. It is yet to offer a way to individuate systems at scales beyond that of the organism acting in its environment large scale ensembles such as societies or ecosystems that realize free energy bounding dynamics but of course part of the question is do they and so VNE was really a theory that was about the organism and the ecosystem from the perspective and the focus of the organisms dynamics in the immediate environment for example affordances even if those affordances were subtle or relied upon this skill intentionality or like you know learn behavior or sophisticated counterfactuals all these things that especially more recently have been receiving focus. This did not articulate clearly or at least had not been specified specifically that it was linked with collective behavioral or ecological or cultural phenomena which is as much a phenotype it's as much biological as an organism level trait like arm length or GM composition that just isn't a difference between biocultural versus bio in this variational framework it's just different scales different mechanisms but it's not even about categorizing different things it's about how you model them and what they are so the variational neuro ethology was a more clearer downwards projection about how things like behavioral and hormonal and neural time scales developmental time scales were related to organism behavior which is physiology neuro ethology but the eco evo divo was less clear the issue was cog gently articulated by Brian Berg and has spent cure Meyer in their critique of variational neuro ethology they asked also whether the Markov blanket and be formalism leveraged by the V&E to individuate systems adequate for modeling phenomena at the scales beyond that of a single organism like sociocultural dynamics since phenomena at these scales may be too transient or not sufficiently robust to license the mb Markov blanket formalism which is defined in terms of conditional dependencies in weekly mixing random dynamical systems which we'll also get back to but BHK catalyzed this perspective and this is going to be broken up into like a few different sections Markov blankets so open road see Markov blankets are the statistical using their words conditional dependencies which is kind of like saying that it's the opposite of the causal relationships and so it's equivalent to finding the causal relationships through their violation of the independence so finding which variables are insolatable from each other in a sense if you think about what kind of information from a generative model you need to understand which variables are conditionally independent or just sort of straightforwardly correlated that's often related to causal state estimation about the world that Markov blankets are unable to explain slash predict slash control ecosystems and other collective behavioral collective behaviors quote quote even though it's collective behavior all the way down at scale because the Markov blankets are not robust enough to capture transient or historical phenomenon so to go to the paper they only work through one example in the paper which was how this hierarchically mechanistic mind evolutionary systems theory can be used within that in any framework to study the scales of human so so biocultural evolution in humans and there's this table provided with some pretty simple slash just one line equations about how different levels of analysis would be modeled but there was a critique which was this downplays a lot of features of ecology and the formalisms that applies a strongly or strategically bounded systems or certain spatial or temporal time scales is mainly be weekly explanatory or valid at different time scales for a number of reasons which is something that we're going to return to when we talk about the ergodic systems so here's kind of a theory tree the FEP is on the bottom it's sort of the paradigm level foundation because that's where we're going to try to make things emanate from and to but again we could just think about what would be the effect of switching that out to why not from left to right we have active inference Timbergin's four questions evolutionary systems theory and then eco Evo Devo so variational neuro ethology is like Timbergin's four questions which was what we heard the quote about with an evolutionary perspective but a simple at least a sketch of one and then more of an emphasis on active inference and on smaller scales of analysis the variational approach to niche construction via NC was also drawing on evolutionary and the pluralist Timbergin framework but drew a little bit more from the eco Evo Devo and active inference side of it has that orange one and where VE comes into play is with bringing in this synergics idea from the variational approach and also the collective behavioral approach and then just again making a new acronym kind of more specifically combining previous acronyms I think it is worth unpacking it's worth tracing out some of these relationships and just seeing what parts get subsumed into what but VE is just it's just a more high resolution area where mapping can now be performed across these different domains so it's kind of like a mapping domain so here's some of the specific parts where they talk about that kind of a mapping domain in this multi-scale framework active inference is inherently a group activity that is the entire ensemble of nested Markov blankets are bound enslaved and constrained by dynamics at higher scales while the lower microscopic scales furnish the macroscopic states at any given level this construction evens is exactly the same circular causality that underlies synergics however here it is generalized to a recursive hierarchy of scales i.e. the hierarchical composition of blankets of blankets intuitively the dynamics at one scale provide constraints technically they establish probability gradients on dynamics at other scales active inference destroys free energy gradients at each scale under the guidance or control of a generative model at the scale above this guidance is exerted through influences on sensory states where circular causality means that the action of any Markov blanket in an ensemble of Markov blankets could be involved in sensing action or perception depending upon its role at the superordinate scale i.e. it has a sensory active or internal state at the level above so this kind of bays all the way up and down approach that they're proposing that from the level of the organism can go downwards into the cellular the neuro and all this level and through time through the evolutionary through the tinberg tolerance and then also outwards construction in the multi-scale setting an effective state like let's just think of our temperature becomes the expression of an eigen mode of blanket states namely the principal eigenvectors of the Jacobian the rate of change of flow with respect to state these mixtures are formally identical to order parameters and synergetics that reflect the amplitude of slow or unstable stable eigen modes in terms of center manifold theory on two solutions on the slow unstable or center manifold in short the Markov blanket of a system or particle at any scale constitutes an ensemble whose order parameters sub 10 blanket or internal states the scale above so that's like the temperature being above equilibrium with the room is because it's part of a sort of closed or semi-closed regular system that's keeping molecules that are in one part of the room constituting my body a different temperature than the rest of the room note that the constituent microscopic states of an ensemble are always blanket states although their order parameters can be blanket states or internal states at the microscopic scale above so a neuron could be sheathing something from a temperature perspective from one side but receiving information in a different conceptualization of the blanket network this follows from the fact that only the only states that matter are those that influence other blanket states effectively all we are doing here is applying the enslaving principle or center manifold theorem recursively to Markov blankets or Markov blankets imagine you are an employee at an institution where you transact your microscopic affairs with other personnel to self-evidence your prior belief that you are good at your job this would entail responding to corporate or institutional goals that emerge collectively i.e. an implicit generative model at the macroscopic level may be homologous to an internal state of the institution macroscopic level relating only to other employees alternatively you could be working on reception i.e. a sensory state or issuing press releases i.e. an active state and this is actually an image from a paper that team com wrote on 9 9 2020 and one of the figures and it was about how remote teams and a lot of other kinds of groups that might not be traditionally considered remote teams can be explicitly mapped and from systems engineering perspective modeled as this nested Markov blanket framework and so you have for example the team members who have their own digital interfaces their own sensory and active states digitally some of these team members can be augmented or can be bots themselves and they're interacting with certain databases and then they're able to also undertake team operations this is where the variational approach to niche construction bank comes in the ANC exploits the expression expected free energy is expected cost plus ambiguity to propose that agents upload as it were much of the legwork in computing expected free energy to their self tailored constructed environment more precisely it argues that niche construction can be cast as the collective activity by which organisms act on their material environment to create the unambiguous structure which can then be leveraged via active inference like creating books and then people can scan their eyes over lasting changes to the niche capture the fact that environmental cues function as an unambiguous indicator of the affordance of action possibilities like if you're culturally sensitized to know what a book looks like you know what it's for you know that if you scan through the first three pages with your eyes and it's culturally specific whether the first pages are going to be on this side of the book on this side of the book or whether the text will read this way or this way this structure can be cast in terms of epistemic resources that fly the actions that resolve the ambiguity associated with future observations while conforming to the prior preferences of the organism entailed by its generative model so you're in the library you're doing some optimal foraging for a book about a certain topic and you expect that there's a certain structural organization to the books and that if you find something where it's written on the back of the book that you think it's the book it is you expect that the title page is going to conform with that so these are like epistemic resources that are more visible more scannable that's what allows that information to be found and therefore leveraged so that's niche construction in terms of direct epistemic resources but you can also think about in cases like social insects where they're extended niche their nest is something that they pass on potentially even intergenerationally and that the structure and the form of the function are all linked to their collective behavioral algorithms with the architecture influencing their behavioral proclivities and there's some interesting work by Professor Pinscher-Wollman on that in the Harvest Rants then that physical structure becomes a epistemic resource as well in that it becomes part of their extended cognition alright so for this last section section 007 of the live stream my friend Sean had a great question and so for this last section it will be a little bit about ergodic systems and as with you know all other areas of this sort of discussion it's a playful and it's a partial recap of the ideas so if someone knows more I would love to go deeper into some of these ideas I just kind of get it out in a way that hopefully might relate to an informative way to onboard on to thinking about things from an active inference perspective and there's definitely perspectives on YouTube and elsewhere that are a lot more rigorous and comprehensive than what I'm about to go through but this is going to be a couple minutes on linking ecology which is what we're trying to explain we're not just trying to make theory after theory we really want to return to the ecology we could make variational ecology a theory about ecology that allows us to do the things that we want to do in ecology whether that's conservation or just understanding so many other things that ecology can teach us but we're also going to talk about ergodicity about cog ergo sum and of course active inference alright so what Sean asked was are thinking beings or their environments weekly mixing ergodic systems if not do they fulfill some of the assumptions or axioms of the mathematics underlying the FEP so this was kind of a great question by Sean because it really came at a perfect time when I was preparing these slides and if we go back to the critique slide where the variational neurology was being presented remember it was that there was the critique of the VNE that got us the response to that that got us to the VE as far as connecting the organism down with the VNE with the organism up with the niche construction just connecting the two points into a line and calling that variational ecology and the critique was that there was going to be some violation of some of the lower level sensory coherence of say an organism and that because of that it wouldn't that the same formalisms wouldn't apply and so this is someone coming from a more like thermostatistics background and seeing a very similar phrasing a different way and just asking a really interesting related question and as he put it the argument of the free energy principle is two parts the first calls on the lawful dynamics of any weekly mixing aka ergodic system that persists from a changing environment from a single cell organism to a human brain these lawful dynamics suggest that internal states can be interpreted as modeling or predicting external causes of sensory perturbations aka embodying and acting a generative model of adaptive policies and phenotypes in other words if a system exists its internal states encode probabilistic beliefs about external states and so when I read that you know I didn't know it's like is he summing that up as like the insight into the whole distillation of something that he finds deeply true or is he bringing this up as sort of a well that's a tautology, that's a circular logic or that's an argument that holds nowhere so I didn't know but it turns out that there is a paper from 2015 called I am therefore I think so this last little part is going to be about that paper which I hope will be interesting whether you're familiar with these ideas or not but also with some of the mathematics behind the ergodicity and about why there's still significant reason to understand how ergodic systems and non-ergodic systems are related to the FEP so I am therefore I think is kind of related to this I think therefore I am cog ergo sum but of course it's just a meme so here's a few other versions I think therefore I am not here or a comic where it says I think therefore I am and then the other person says I don't therefore I am not and also the cog cogido ergo sum I think therefore I am I think therefore I overthink I think therefore I regret I think therefore someday I won't so all these different implications what are the real implications what logic can we make claims about well that's not exactly what this is about this is going to be about two kinds of systems which are ergodic systems or not or ergodic ergodic systems and this isn't something that's a hyper abstract but it definitely does recall a lot of generalizable principles and math so to make it very specific and empirical we're going to discuss it in terms of the ergodic hypothesis so it's a specific hypothesis that we're talking about that means it's something we're going to test with data and the hypothesis is we hypothesize that mixing through space is like mixing through time and let's talk a little bit more about that because it's kind of a wild and interesting idea so the first thing to note is that it's a hypothesis this ergodic hypothesis it's not the ergodic law or the ergodic pattern or anything like that it's a hypothesis that can be tested using data and statistical models that are transparent and reproducible and just open to understanding how changes in their parameters influence the kind of conclusions that are reached so that's something about what we're discussing here with the ergodic systems also the ergodic hypothesis depends on the spatial and temporal scale and then also the measurement or the analysis space and potentially this comes into play a lot more when you're really getting down into the weeds of the math or the application but often the way that you measure something or which sections of the time series you allow for burn in for example or mixing these things are all really critical and what scale whether you're looking at a tiny tiny tiny cube and seeing if there's random association between chemical particles for example versus looking at a one meter cubed blocking of the world so the coarse graining matters a lot now ecology and history are often strongly critical of applications of ergodic methods or that methods that often are used to model ergodic systems so first let's talk about the ergodic hypothesis and what it actually says and does and then talk about some ecological critique of it and then see how the cogego sum comes into play alright so the ergodic hypothesis is defined by the time average being equal to the space average so that is again this sort of mixing through space is like mixing through time idea and one way to think about it is like if you have a box with a gas in it and you're going to trace one molecule and take snapshots of it at a hundred times would its trajectory in some statistical way be similar to if you just selected a hundred particles in the box at the same time or if you let loose one molecule bounce a hundred times by itself all these different ways of asking it but it kind of relates like is it basically fair to sample one roulette table through time or a hundred roulette tables at the same time every roulette spin is like the same so in it's very different than something where the strategies evolve through time for example you might already even be seeing and one more slide of just the definition is that there's this generation of an ensemble so that's sort of already for saging this ensemble group selection dynamic versus the time element and this paper was very interesting that I found while preparing for this stream and I'll just go through it because I think it's also a really cool story about philosophy of science and about how the way that we think about different hypotheses and axioms and principles really influences the kind of science that we do so it's called from the ergodic hypothesis and physics to the ergodic axiom in economics like Kirsten 2015 so the work is summarized as this contribution seeks to clarify the specific relation between the idea of non-ergodicity which is drawn from statistical mechanics and its role in and for economics and finance the identified change of status is how the ergodic hypothesis and physics became the ergodic axiom in economics therefore we follow the idea from rational mechanics of Gibbs via Wilson to Samuelson and eventually from the latter into rational economics the methodological spillover from the natural sciences to economics and the assumption of ergodicity in particular enabled and shaped the mathematicization of economics seen since the 1940s so it shapes the fields direction depending on how people use methods from ergodic assumptions or not and this was a nice diagram where Boltzmann and Poincaré and Gibbs and how it transduces through different generations and it's kind of like a citation network here and I think a great example which was used in the paper is the St. Petersburg Paradox and this is a really fun kind of game theory and thought experiment complexity classic that's related to economics so I'll just read it as it is from the paper in 7 it's called the St. Petersburg Paradox in 1713 Swiss mathematician Nicolas A. Bernoulli posed in an exchange of letters with a French mathematician Pierre Remoudre de Montmore the following problem what is the reasonable price for our ticket for the following lottery so on the first flip it's going to be a fair coin and if you win you can either take one dollar and then games over or you can play again and then on the second flip if you keep playing there's multiple versions in this version in here it either gives you a dollar or you keep playing and so you get a payoff at each time if another version is where you can decide to keep playing you can imagine getting a 50% chance that first one is kind of like now the games are very similar because you can either by force or by not so in either case you can ask yourself how much would you pay to play either of those two little variants of the game and it turns out that people don't want to pay infinite money to buy in so especially the one where you either continue or get the payout the expected value of all the ensemble of trajectories one in a million is going to make two to the end money and so one in a million is going to make so much money so the expected value across the ensemble with the geometric mean or the arithmetic mean would be like very high but it still wouldn't be worth that much why so let's go into that why does that St. Petersburg Petersburg paradox challenge ergodicity one answer is it's about pragmatics your life is not ergodic you're not well mixing across you know different places to eat across different places to sleep there's a regularity to your life and so especially considering it in the sort of niche construction for your own uncertainty reduction precision maximization even putting that all aside it's just your own singular trajectory from a strategic perspective doesn't resemble the assumptions that are under the distribution conforming norms of well mixing gas molecules that's for sure and another way to say that is that the realized risk profile through time is not the same as through space so if you're playing that game who wants to be a millionaire and it's going up okay I've got 12 coin flips in a row I'm at a thousand dollars or something like that I could take a thousand dollars or I can go home with nothing at some point it's just not going to make sense to go double or nothing on some insane value because there's this realized tradeoff between risk so that's not and that's part of this historicity and non-ergodicity of your own trajectory ecosystems may not be ergodic so that would be like because of secession dynamics or nonlinear perturbations or resilience properties ecosystems may or may not have this sort of roulette wheel property or other nice aspects of high low dimensional systems that physics prefers so there's interesting theoretical work oh well imagine if every patch of the ecosystem was like a spin blast that would be a phase transition but in the real world it's always a little bit messier than that so then just take that sort of saint petersburg paradox and literally the money one now adding each construction and so I was just thinking well yeah what if what if every city in the country gets to play this paradox and then one city or you know one individual wins just a ridiculous amount and a few people win a small amount and most people win nothing and then that city does infrastructure investment so how does that change the way that future people would be able to for example play the paradox later so that's niche construction so now think about the historicity and the niche construction whether that's social insects or humans and you can see how some of these ergodic assumptions as they apply to cultural behavioral systems maybe like a little bit tenuous or metaphorical at best here's a few nice figures from this paper though too so on the bottom right there's this ergodic case and so one of the features of ergodicity is this time reversal like the molecule bouncing around the room it's kind of this general time reversible family of models that allows extremely interesting bi-directional computation in a way that historical systems don't and so timelessness exists in the current T equals 0 that's the present moment and then there's like this analytical time like reaction coordinate for a chemical reaction when the two molecules come together it's not graphed in terms of time but reactions do have time distributions that they occur on but the way that that is represented is on actually free energy landscape Gibbs free energy with a reaction coordinate because that's like this ergodic hypothesis about the way that these reactions are reversible through time in contrast there's the non ergodic case on the top and in the top there's a symmetry break in the present moment so that means that there's events in the present moment that change the distribution going forward in a way that doesn't make time reversibility attainable proposition and that's because they happen in historical time so that's kind of a cool idea and then there was yet another great figure from this paper where with kind of a lot going on but I think it's worth looking at because it was so cool so first there's the not analyzed not analyzed is just before your data windows available then there's this realm where you use the ergodic methods ergodicity is almost like the name of a city and what you see in this window is the spurious quote causal relationship in your observation window so let's just say like you know between bitcoin and and exports of some other resource or something and even if you everyone knows correlation isn't causation but then there's methods like dynamic causal modeling danger causality and machine learning methods that try to quote learn so there's so much in the way that people use methods whether or not they think they're they at the very least will say they're using data from the past reduce their uncertainty about the future which is what led to financial collapse it's what it's what leads to a lot of strategic strengths and weaknesses in a lot of different areas so that's the ergodic fallacy valley now the non-causal window is like let's just say where there's a different window of time whether it's window two or window four and under either of those time intervals under consideration the same factors that were at play and one may not matter that much like if you observe the first post forest fire three weeks the rules that dictate that regime let's just say the spurious causal relationship oh nitrogen is limiting factor or insects don't play a major role in this process that may or may not be true in a future time window that's not overlapping or partially overlapping seems pretty fair pretty valid I think it's a great figure and then also there's like second and future symmetry breakings as well so even if you go well there was a major systems change and the present moment or at X time but things are still the same in this regard this invariance still holds or this measure is still valid but then they they can't guarantee that there's going to be this invariance in a future arbitrary symmetry break so what are the frameworks that will carry us through these symmetry breaks and not just have very weak underbellies with respect to things that we don't even perceive on the horizon so let's return to Debra's paper about fusion of behavioral ecology and ecology and talk about whether simulations variability historicity and other sort of not quantitative questions but qualitative ideas challenges ergodic hypothesis so she wrote in behavioral behavior in the illusion of ergodicity in ecology so section this paper ecology often represents ecological processes in terms of the functional relation among average values for example to ask whether an invasion will be more successful when there are many species in the community to ask how an average property will affair when the average population size of many local species are high methods for evaluating ecological outcomes often assume ergodicity that all samples are homogenous and that populations being sampled are invariant this can be misleading behavior is a response to conditions but individuals vary in response and conditions change the result is that ecological processes are not ergodic samples are not homogenous and processes can change over time very nice and a well titled section and to kind of add on to this perspective and return to this question about non ergodicity in historical processes like ecology here's the paper in 2016 replaying life's tape simulation metaphors and historicity in Steven J. Gould's field life so not a endorsement or critique of Steven J. Gould it's to bring it up as a philosophy of science point just as any other paper I would address so this is from the paper and it is a 1974 paper of Ralph and Gould and what they did is they did simulations with where branches node branching rates were a parameter and then also there was like mutational changes and they're like wow the extinction distribution looks like we see in the fossil record and the morpho distribution looks like we see in the fossil record maybe selection doesn't matter or maybe it's just random stuff and to kind of build on that in another figure in the paper this is two outcomes of the simulation that they later ran see this is probably a later paper from them or from colleagues that use again kind of like a birth death process so like the lateral line is a birth of like a new clade speciation event with a of some type maybe with niche displacement maybe not and then there's also the population size growing or shrinking by a small amount but when it reaches zero it kind of goes extinct and so they were seeing wow we're seeing big variation between quote clades or different levels of organization in terms of their biodiversity or in terms of their species number or their morphological diversity so again maybe it's just a random punctuated equilibrium events that end up driving the ship now this comes down and you can come down on the Steve and Jake Bulliside or not on this question whether the historicity of the past the uniqueness of the trajectory of the timeline that we appear to be in whatever your feelings on the historicity of that the question is really from an instrumentalist point of view from a functional point of view can you use these methods that may have statistical assumptions underlying them that assume ergodic processes so for example when you do a t-test it's like making the assumption that with respect to other variables all else is equal but and everyone says of course if you sample group A in country one group B in country two it's conflated but the t-test doesn't know that and so where we don't know that the t-test doesn't know that and more complicated machine learning methods are only able to do this in a really more nuanced way in certain situations especially with large data sets so for the Steve and Jake Gould camp it was the fundamental long-term deep time events blocks one events that drove the ship whereas another way to think about it in the one that I personally would probably end up emphasizing at the end of the day is that it's actually selection on the morphology locally over that deep ecological time there's also the day to day of those little trilobite creatures having to live in their niche and so selection is continually on their morphology and yes events happen over longer time periods and the whole question about multi-scale biologies how to integrate those two rather than saying well oh well clade formation is random versus oh it's really strongly driven by niche partitioning so to kind of actually close this with a with a cogito ergo sum part here's some figures from the I think or I am therefore I think paper with Friston and Mathias so this left side is I am ergodic therefore I think let's just you know see what the authors say because with all these critiques from ecology and from cultural biologists on ergodicity and on the brain behind everything so here is sort of a two-stroke engine and it goes from resis extensa which is sort of the the behavioral action policy selection flowing out and then the rescock unit is like the belief partition of the brain so the belief could be I believe that the train is going to come in five minutes and that would be like corresponding to a detachable thought or debate whether it's a representation or whether it's a metaphysics component and then this is just one partitioning they're thrown out there and we'll see another one in a second and then this free energy functionals being used to select the maximum likelihood policy and so that's why it's doing this pursuit of the minimally surprising across the variational scales of analysis of surprise is trying to select the control policy that's going to bring sensory observations into alignment with a deep generative model so the belief I am ergodic could lead to behavior that at least pursues ergodicity is that enough and then they kind of build it out one another level which is I think I am ergodic therefore I am so that kind of is an inversion of I am therefore I think I don't know how you get from here to there maybe there's other figures to make or other arrows to draw but it is interesting to think about how basically it's a little bit unpacking the Res cognitive beliefs from a Bayesian perspective into the posterior belief and then the likelihood and prior belief so that's like the now there's like a Bayesian two stroke engine up there and then out of the free energy functional reflecting the likelihood and Bayesian beliefs is also the policy that sort of relates to our discussions on the generative model being a policy model alright here's kind of a funny figure with focusing on the ear of this statue and just it has equations on the ocular motor being linked through the action so action causes the ocular motor to move and the eye sends back visual sensations and proprioception so that's SQ and SP and it's like that's the interface for the eye that's the API for the eye that's what we have access to and that's what we can model and it's that dual instrumentalism one layer is like yeah these are the areas of signal that the eye gets and receives so from an engineering perspective the scientists instrumentalism is like yeah it's good enough for us just to model the system that way because there's basically nothing that goes outside of that flow state even though each one is complicated but then the second one would be this is actually how the eye works and it turns out that some of these hypotheses about how the actual streams of information or communication should be in the brain that they are concurrent with some of these inference schemes which is interesting and this figure was about how like looking at the same statue there's the stimulus which is the actual image itself and what the visual input is is that is the retinal cells which are activated let's say by edges and then the salience could be in terms of a deep generative model as far as what area of the sample would be what area of the visual field would be most optimal to sample ergo which action policy would be optimal to undertake and we see this sampling which is a little bit outlining the face if you kind of look at it and overlay it because these are the optimal places to forage so to speak for information to resolve what it is so I'm just going to go through this last figure from I am there for I think because I think it's kind of interesting funny so here's how we can walk through this a behavioral sequence is embodied so that's the ocular motor red dots moving around it's like going so the behavioral sequence is embodied or enacted and also encultured and embedded that's the whole variational approach so right now we're focused on the ocular motor dimension it's not really unpack it but it's something that we could have moving up like in one culture they've learned that avoiding eye contact is you know or people's status is signaled on this part of their body or no it's in the eyes where you determine something so this is not just a one size fits all solution for human vision this is about a model framework that we could use to model what we want to model but a behavioral sequence is embodied by acting to converge towards beliefs about hidden ocular motor states so right down the y-axis is this hidden ocular motor state and then there's the proprioceptive expectations so that's like if your muscle with one side like if your bicep contracts your tricep will relax and if that breaks down if that communication of expectations of oh hey I'm contracting or I'm contracted if that breaks down you get a lot of different types of muscle phenomena ranging from you know pulled muscles and strains to like post injury or even without injury changes in the muscles perceived healthy range of motion or limit or minimum or maximum force and that has to do with a lot of the proprioceptive information being integrated with beliefs about the hidden states what that results in is that observed sensory states which these are the visual samples so that's we're taking one the hidden ocular motor states we converge towards belief on those what that results in is the observed sensory states so here's like an eye and eye and you know there's five dots here and so you're gonna get more sampling on the eye but that's kind of what the camera sees and then it says absolute data or surprise or prediction error because this here is like showing the visual samples as kind of camera shots of what is being seen but in the predictive processing framework you could imagine that actually what's coming back is like a parameter update message passing thing that's actually about deep expectations about what to see so that's not even included in this model but that's part of other models and what that does when the sampling is like I expected an eye there I looked at her you know third eye and I didn't see it visually you know hypothesis confirmed so that type of sampling is also informative and so what we see on this graph is that the expectations about hidden perceptual states so our sort of fuzziness on the whole map decreases through time and what that results in is that the corresponding percept remains sharp and concurrent with a coherent generative model through time so I always just think about vision in this case like your whole visual field unless there's other secondary issues is clear and has color vision across the entire field and so this is maintained even though there's a large blind spot your resolution is lower outside of your central area and also there's not as much color perception outside there's a bunch of other visual illusions too so it's clearly we're perceiving our generative model visually and why not in other areas too when it's as clear or clearer so we have this face just coming into clarity based upon a deep generative prior which is why when there's things that you don't know how to extract the information from it's hard to know how to do ocular motor scanning let alone to how to do the semantic mapping now isn't that the big question is will this kind of static face fixed image scanning to resolve uncertainty will that level of functional closure relate into eco-evo skills of analysis so you can imagine this is just a static face doing classification but would this extend into adversarial relationships with one person would this extend into deep counterfactuals with valence and skills and team information hypotheses all these things playing out that are happening in social scenarios would that be able to play out this is just eye scanning so will that level of analysis of the group or the community the evolutionary trajectory the cultural unit the ecosystem any ecosystem will that level of analysis lend itself to this multi-level framework and isn't that the question so that is all I thought I would close with it as an opening to whether people thought yes or no or I think I'd like to know this piece of information before we made a decision I'm open you know I want to learn more I want to figure out what that actual answer is is it going to be that we'll have some very tight and coherent markoff blankets that we can just do simulations and understand the interactions between and then let that higher level parameter just freely or will it really be the case that there are order parameters that are just beyond what we can currently understand as mapping onto this framework or some other framework who knows anyways that was kind of an interesting discussion so I hope you liked it and I just say thanks for participating we will provide a follow-up form to live participants feedback suggestions and questions always welcome and yeah just stay in communication we're going to have number 7.1 and 7.2 in the coming two weeks and if you want to participate and you're listening before them then get in touch also we'll hopefully still be looking for papers to read and for people to have on discussions as long as you're listening to this so just give us a comment or a message and it'd be good to get in touch so any thoughts especially on this video on where does the FEP fit into all this where does active inference fit into all this what are we looking to understand with variational ecology how are we going to use variational ecology how is it going to map onto classical ecology what ecological questions are we going to be able to address first how is this going to relate to the kinds of explanations and design solutions that ecologists have worked with before versus proposed deeply hypothetical systems a lot of interesting questions plus of course hearing from everyone else on their perspective on the implications or strengths or weaknesses so anyways cool thanks for listening I hope that you had a good time and found some parts of it interesting and I will talk to you later