 Hello and welcome, everyone. This is the Active Inference livestream. It is September 29th, 2020, and this is our episode 5.1. Welcome to TeamCom, whether you are a first-time listener, participant, or whether this is not your first time. We are an experiment in online team communication and learning related to Active Inference. You can find us on Twitter at inferenceactive. You can find us on email at activeinferenceatgmail.com, through our public Keybase team, or through our YouTube channel. This is a recorded and archived livestream, so please provide us with feedback so that we can improve our work. All backgrounds and perspectives are welcome here. We really want to hear about everyone's individual experience. And also, as far as video etiquette goes, please mute if it's noisy in your background and remember to raise your hand. It's awesome that we have so many participants today. And so, to hear from everybody, let's just make sure to raise hand and all keep a cue of people who have raised their hand. So, today in Act In Stream 5.1, the roadmap for our discussion over the next hour and a half will be as follows. First, we'll have a warm-up section where we give a quick intro and check-in. Then, we'll be discussing the paper. We'll talk about the goal of the paper. We'll talk about the roadmap or the section headers. Go through the abstract, and then go through all the figures. We'll also be able to touch on some background concepts if we have time and address some excellent questions submitted by Sasha and other participants. And next week on 5.2, we're going to have further discussion on this same paper. So, please remember to save and submit any questions that you might have that don't get addressed in this first discussion. Awesome. So, here we go into the introduction section. Please introduce yourself and your location. And just say hello to everyone in a short introduction, and then pass to somebody who you haven't heard introduce themselves yet. So, I'm Daniel Friedman. I'm in Davis, California. And I would like to pass it first to Alejandra. You're muted. Yeah, well, hello, everybody. Well, I'm in Mexico, specifically in Cuernavaca. And this is the state where I'm living now. But, well, in my professional background, I'm a full-time associate professor in cognitive and behavioral sciences at the School of Psychology at UNAM. And also, I'm an active member of the cognitive robotics lab at the YM Mexico, also. And, well, my research focuses on perceptual optimization processes, such as sensor attenuation and sensor enhancement. And also, the computational modeling of perception, cognition. And I'm, like, hooked with the predictive processing of the active inference by decoding the energy principles since a long time ago. It's like my passion. And I'm, like, giving a full-time, some of my days of reading and trying to understand this particularly free energy principle. So, yeah, that's me. Let's pass to Alex. Hello, everybody. Yeah, my name is Alex. I'm from Moscow, Russia. I'm a researcher in system management school and trying to find a way how to join concepts from system engineering and active inference frameworks. So let's say for Sasha. Hi, thank you. My name is Sasha. I'm a neuroscience graduate student at Davis. And my background in free energy principle is just reading and enthusiasm for figuring out how to apply this to my own work with developmental neuroscience. And just looking forward to the conversation. I'm going to pass it to Mel. Hello. Yep. Can you hear me? Awesome. Perfect. Hi, I'm Mel Andrews. I'm in Cincinnati, Ohio right now. I have a background in copy of science and always been fascinated by issues of theoretical biology and switched to philosophy from my PhD to really focus on sort of what I called life-mind continuity theories. And in particular, I'm focusing on the free energy principle as a formal articulation of a life-mind continuity theory. Oh, I'm going to pass this on to Richard. Thanks. So my name is Richard Niemeyer. I'm an assistant professor of sociology in Denver, Colorado in the United States. My interest is mainly in understanding how we can integrate neuroscience, cognitive psychology, and sociology into a more integrated framework. Most of sociological theories drawing on a lot of behaviorism from the 70s and 80s is kind of its theory of mind and how the mind society relates to one another. So I'm trying to figure out more up-to-date ways to integrate mind, body, and society. Most of the mathematics that I'm familiar with in terms of kind of how to do that is a lot of fractal geometry, graph theory, dynamic systems theory, which is really fascinating. But I've been really interested in kind of the free energy principle. It seems to do a much more kind of intuitive, fluid, dynamic way of integrating different levels of analysis and integrating specifically cognitive emotion, action, things like that, which are really critical when you get into that kind of social level of reality. So I've tried my best to work my way through a lot of papers by myself and it's been a struggle. So to find a community who's kind of doing this together and can kind of listen to other people talk about these ideas is really, really helpful. So I really appreciate the opportunity to be here. I'm gonna pass it on to Muddy. Oh, classic, it was on me. Hi guys, my name's Muddy. So I was a PhD student with Carl Briston graduating back in 2016, but I also was a medical student on the MD-PhD pathway. So my interest was in applying free energy principle to study stroke dynamics noninvasively using neuroimaging. I since left the medical world, I practiced medicine for about a week and then went into the tech world, I worked with Google DeepMind and started the medical tech company, which was acquired by the company I currently work with now versus where I'm the director of AI. And we're really interested in spatial web technologies, spatial web protocols, and building provenance AIs that are not only explainable, but action-based and adaptable. So that's kind of my background. And I will pass it on to Professor Klippinger who is quite an acceptor. Thank you. Hi everybody, my name is John Klippinger. I am now in Northern New Hampshire, Jefferson, New Hampshire farm, but actually I'm generally out of Cambridge, Mass. And with the MIT Media Lab, the city science group, and I've been at MIT, the human dynamics group for a number of years. My interest, actually, I'm a full convert, recent convert to the Friston cult as it were, but I continue to be amazed by a continued reading and the papers that are coming out and the productivity of this. And working with Maxwell and Casper about how to apply this to modeling new concepts of firm, new concepts of market, new concepts of community. So we have some projects that we're focusing on that and actually collaborating with Muddy on that as well. Cool, let's go to Shannon. Hi, I'm Shannon Brooks. I'm in South Dakota currently, but I'm usually based at the University of California in Merced. And I'm in a sensory motor neuroscience lab. I investigate the neural dynamics of music cognition and specifically how we process rhythm and then interested in how predictive processing approaches can help us understand how we perceive rhythm and music. I'm also interested in how groups of people interact when they're experiencing music, when they're playing it or listening to it and in larger dynamics of crowds. I'm excited for today's discussion specifically for how a crowd could have a different markup blanket based on its interactions throughout maybe a musical performance or a sporting event. Cool, I think we have Maxwell. I'm muted Maxwell. Sorry about that. Hey, I'm Maxwell Ramstead. I'm a postdoctoral fellow at McGill University and I'm one of the main architects of the multi-scale active inference formulation that we'll be discussing today and the first author on this paper, multi-scale integration beyond externalism and beyond internalism and externalism. And I will pass it off to Alex Kiefer. Hi everyone, I'm Alex, as Maxwell just said. I'm coming from a philosophy in cognitive science background. I got my PhD at the CUNY Graduate Center currently at Monash University in Melbourne but not actually there because of the pandemic. So I'm in the New York City area and I'm interested in, I'm coming from sort of a traditional philosophy from my perspective. I increasingly got into like connections and machine learning, modeling, eventually the FEP and I'm interested in how this stuff ties into fairly traditional representationalism on the one hand and sort of physics on the other and how these come together. So that's my angle on this. And I'm really excited to talk about this paper. Cool, I think Steven, and then if Daniela jumps back in, we'll go to her. But Steven, I think you're the last one here. Okay, hi, I'm Steven, I'm in Toronto and I do community development projects using theater and immersive participatory approaches and ended up developing a process called socio-drama topography in South Africa in 2006 and China scale that process in different ways has looked at how embodied cognition fits into this kind of process of engaging communities to make embodied maps together around issues. And that's kind of translated over time into other sort of immersive practices. And I was looking at embodied cognition over a few years and I kept finally seeing this work with Carl Friston. And I thought, oh my goodness, I've already gone so far down all these avenues and I've seen this other thing. And then suddenly I found a page which spoke to how many citations he had. And I thought, wait a sec. And I suddenly saw that his work was actually bringing all these different threads together into this unified framework. So, yeah, I'm really excited by that. Awesome, well, thanks so much everyone for coming out to participate. It's really a spread of countries and levels of experience and fields of expertise. So it's about bringing it all together and coming together around these topics. So just as a warm up and to hear from everyone who would like to speak by raising their hand, let's just start with asking, what drew you to this paper or topic today? So I'll start and while other people are raising their hand and I described this a little bit in the 5.0 video, but I think the idea of integrating multi-scale systems is really interesting and there's so many ways to go about it. And it's almost undeniable that we're living in multi-scale systems, but the question is, what now? And so I'm always interested to hear about systems for thinking about what now and also for different people's perspective. Great, Muddy and anyone else can raise their hand. Muted though with a video off. It's the fine line between muting the background noise and being ready. So I understand, but go ahead, Muddy. My apologies. Yeah, I think nested systems are an incredibly important way to view real dynamical systems that we see all the time in the spatial web. So we constantly have streams of data coming from IoT and then we're kind of met with this scalability challenge of how do we actually model a system accurately to represent the dynamics that we wish without getting bogged down in this sort of field of complexity? So by abstracting away these layers into really, really meaningful nested heterarchies is really, really key in being able to practically deploy some of these novel learning paradigms in real-world spaces. So I think without this sort of quote-unquote dimensionality reduction, and I mean that in more of a philosophical sense than a mathematical sense, we're going to struggle to apply AI at scale. So that's why I'm particularly interested in the work that Maxwell and the free energy principle group is doing. Cool, let's hear from Maxwell, then Sasha, and then I also see other people, just those two next. Maxwell also muted. Sasha, do you want to go first? I don't have to go next necessarily. Sure, I'll briefly say a bit. I was drawn to this paper because as a neuroscientist, I think a lot about multi-scale levels and to decide at which level a system should be studied because even though we're setting something at a molecular level or even in a different species, we ultimately want to make claims about larger cognitive processes and how to do that without abstracting away really important aspects of the system at the scale that you're studying it. It's a huge question. So I really like thinking about these Markov blankets at different levels and where to draw the line. Cool, Maxwell, then Richard, then Alex, Kaye. I think my remarks are going to echo a lot of what's been said. I got into the free energy principle stuff originally because I was looking for a theory that could address dynamics and across multiple different... Well, the way I used to see it is ontological levels. In working with the free energy principle, I've come to reject the traditional metaphysics of ontological levels in favor of this view that there are big things and there are small things and big things are made up by small things and big things tend to be slow and small things tend to be fast. And what you actually have is like this nested series of systems that are made up of increasingly big things. And yeah, back when I was starting up my PhD in 2015, there wasn't any theory that I could see that was really able to not only define dynamics at each of these different levels but propose a way, a mechanism by which they were all integrated. And this is what I found via the free energy principle. So I sort of made it my thing from 2016 to now actually to explore the multi-scale implications of this framework. Awesome, Richard, then Alex Kaye. Yeah, I think one thing that I've been really drawn to the last couple of years is the mechanistic philosophical perspective of Carl Craver and Lindsey Darden. They give you this really nice, I think, tool set on how to talk about decomposing mechanisms, levels of mechanisms, which is very intuitive and I think works well with sociology and biology, which has this very clear conceptual way to move up and down, quote-unquote, levels of reality. But the thing that's I think always been missing is how do I take that conceptual language and match it with an empirical framework that allows me to model those levels of reality in a way without losing any of the subtleties and any of the conceptual apparatus that is so critical to talking about those different frameworks. And the more and more I get into the free energy principle, it seems like it can give you that empirical language that has a very nice natural transition into that conceptual language I'm so used to that gives you all of the stuff you're trying to talk about conceptually, but then also gives you these empirical tools that allows you to go above that and talk about the organization of the mechanism in a very nice, precise, empirical language that's just impossible to do, I think, with a lot of existing methods in the fields that I'm used to. Cool, let's do Alex, Kay, and Steven, and then I'm gonna go to the second question. So I just recently sort of got into this whole multi-scale picture associated with the FPP, and I guess my interest in this is it seems to me like I'm nearly 100% convinced that this is a really good description of reality at multiple scales, but I'm not completely convinced that this has the radical kind of philosophical implications that some people think it does. I mean, this paper actually strikes kind of a nice middle ground, but I guess to put it briefly, like I view this as we've got nested multiple systems each of which possibly has a physical as well as a cognitive description and maybe the bounds of cognition within that hierarchy or maybe less strict than may have been thought. Like I'm happy describing cognition or us grabbing cognition to systems at various levels within this hierarchy, but I don't see that it's so hard to pin the cognitive properties of one particular system onto a particular physical system. So I'm not sure, anyway, I hope that that gets, we can get through some of that. Also, I'm not sure which system I am in this hierarchy of embedded systems that, you know, my brain involved many systems in itself and that's itself embedded in larger systems. So I'm trying to figure out which type of system within that hierarchy an individual person might be identified with. Cool, Steven, and then we're gonna go to the next question to Muddy and Mal or anyone else who raises their hand. Yeah, I'm very interested in how this gives this kind of action orientated MISO level approach, which is really useful, because when I'm working with their community, trying to understand how to connect to the individual, to the group, to the team, to the organization or maybe influencing change of the municipality or broader social levels with this real problem of bridging the scale. And it ends up getting caught up in all this language which ever field has been, you know, the psychological perspective or this perspective or that perspective. So this gives a kind of, seems like a very neutral, clean way of understanding how people make sense of different levels. And that's really valuable. I think that's been missing from community psychology is really actually psychology itself suffering from that because it's kind of stuck in this level of perspective or interpretation. So this gives a way to actually go into the sub-personal. So for me, getting into all of this has actually opened up all this sub-personal neuroscience side which I actually haven't really got into before. So it's kind of interesting. Nice, thanks for all these responses. Mal also raised her hand. Yes, so next on the stack is Muddy and then Mal and feel free to answer either the first question or the second question about what a multi-scale system you work with is. So Muddy first. Great. So they actually kind of fold into each other. So one of the things that I wanted to mention was the practical challenges of implementing inference in industry, right? Which we kind of do it in a black box way. I think it was Sasha who mentioned that or Alex who mentioned he had an interest in this as well. And what we find is that the practical correlates in terms of hardware and chip design that we need to deploy this stuff in reality is difficult based on the data sets that we collect. So one of the systems that we're currently modeling right now is warehouse pick and pack working for Amazon. And understanding that the systems that drive why people pick certain boxes from certain shelves, for example, is like a multi nested system of everything from how valuable an item is to how tired the pick and pack worker is, even down to the function of how tall they are versus how high on the shelf the particular item on the waiting list is. So understanding how to pick that apart and deliver value essentially in industry eventually. So yes, we have this wonderful way of viewing the world, but we also have a practical set of tools which are a wonderful way of modeling the world as well. And so having this nested system, I think is really, really important in terms of biting into the problem of practical deployment of these types of models at scale in order to deliver value. Now, I don't just mean business value, I mean, you know, inside value as well. So those are just a couple of comments that I wanted to introduce into the discussion as well. Thanks for that, Mel, and then anyone else can raise their hand. Hi, I was drawn to this topic because I've been trying to convince Maxwell for a long time that the FEP is just, that basically we should be taking as instrumentalist approach to this as we possibly can. This is like the bare bones. It's just a model. It doesn't constitute knowledge of the world. It doesn't constitute falsifiable claims. It doesn't constitute et cetera, et cetera, et cetera. It's just a model. And we use it to get at things in the world and that's it. And this paper is interesting to me because in it, Maxwell is the one pushing the instrumentalist line that I don't wanna write. Namely, that there's nothing more to the boundaries of systems than is it useful to model it that way, right? Can it be used to be modeled as the debris is here? For me, I'm very willing to privilege the organism as a privileged scale or level or unit of analysis and I'm very willing to privilege the brain as well. In that order. Interesting, fighting words, Mel. Shannon, and then anyone else can raise their hand too. I feel partial to those last words from Mel on privileging the organism or the brain. But that could be just because I study the brain. I wanna know the neural mechanisms that are giving us any of our perceptions or engagement with the world. But I think I was drawn to this topic because I was drawn through predictive processing literature and like Andy Clark's work and then eventually on here to the free energy principle. But a lot of what was written and like is now has these great mathematical models that we've touched on before, like an ability to translate maybe some sort of phenomenological experience into a formal model. It meshed really well with my studies in music and what like teachers of music will instill upon you in that you're basically through your musical experience, you're modeling something about the human experience. So you're choreographing aspects of tension and release which sounds a little bit like balancing predictions and prediction errors. And you're trying to maintain a certain boundary within a musical piece of the emotion that you are experiencing or that you want your audience to experience. And this is sounds a lot like a boundary on a certain phenotype of experience. And just sort of this like superficial matcheting of the language of musical expression and then this empirical like statistical language is sort of what drew me in. And I suppose that answers the multi-scale system I work with music and musicians and people. Cool, very interesting. We'll go to Richard and then the third question. Yeah, kind of a building a little bit on Mel's point in terms of the multi-scale system. So obviously as a sociologist, I deal with groups, I deal with society. And I can take an individual and I can take two individuals. Now I have a dyad, I have another one, I have a triad. I can model that as a network. I can collapse the network down and model it as a dynamical system. And now I can talk about organizations. I have multiple dynamical systems, talk about them interacting. Now I have a network of coupled oscillators. I can go in, I can use fractal geometry to talk about how properties are scaling up and down from the dyad to the triad all the way up to that network of coupled oscillators. And at the end of the day, I'm never really confused about whether or not I'm really talking about networks or am I really talking about dynamical systems? They're all just models that allow me to capture the phenomenon that I do believe is real. I do believe society exists. If you've ever wanna get into a very strange conversation, go to a conference of sociologists arguing about whether or not the United States is real. And my comeback is always don't pay your taxes on April 15th. And you'll figure out very quickly if the United States is real or not. And so I think a lot of these philosophical questions are always fascinating. Especially at three o'clock in the morning in a pub when you've had too many to drink. But I think ultimately at the end of the day, there is a concrete reality. The question is what empirical model just allows you to model it as true to the theory or the concepts that you're dealing with without losing information unnecessarily. So I think, again, I think any empirical model that allows me to do that is fascinating. But at the end of the day, I do believe the humans that I'm ultimately talking about are concrete in our meal. Nice. I'm gonna introduce the third question. We'll go to Buddy first here and anyone else can raise their hand. The third question is, why does our philosophy of systems matter? That is quite bizarrely scary and psychic because that's what the point I was gonna make just to piggyback on Richard's comments just there. Which is, I remember these discussions, these really, really lively discussions about how representative is this model of reality. And I remember at some point just thinking, well, how useful is this a question to ask? At some point, is the model just a tool, a means to an end? And do we employ bounds on what we expect from these models in terms of their utility? And if these models and their outputs satisfy these bounds, great. Sometimes I see some of the discourse around free energy principles and what it means philosophically. And we lose what it gives us in terms of explainability and actual measurable dynamics. So I actually fall probably like very far on the side of I don't think our philosophy of systems matter at all. I think our explainability of empirically measured data is by far the most important thing and trying to explain systems where our data sets are incredibly sparse. So for example, the entire hierarchical organization of the brain from cell to effective connectivity, I think is sometimes missing the point of just how valuable these novel schemas can be. So that's kind of what I wanted to touch on that. Nice, Steven, and then anyone else before we close out this question. Yeah, I think that that empirical pieces and practicality or practice application is really a good point. One thing that has caught me out though, and I've realized is that perspectivism is very much part of systems generally. And it's also part of psychology, like we take a perspective on stuff and there's this assumption that everything, well, what's your perspective? And that word just gets used all the time. And it's like, but the thing is that the free energy principle gives you a way and then they say, okay, and now we're gonna do stuff that's performative or arts-based. And it's a bit like, we don't know what to do with that. So we just call it another field, right? And then they do that. But the active inference gives you a realized, well, there's a way that you can abductively or from within be creating meaning during experience. And it's, I suppose, a bit like that person picking stuff from the shelves. And they're not sitting there as they do it, having a perspective. Like they may take a perspective if you give them a qualitative survey later, which is what nearly all social science does. So you take a perspective, a first person perspective on your experience, but the experience wasn't first person perspective. The experience was just the experience. So this gives a way for that to all be kind of unifying in some way. Cool, let's go to Alex and then Richard then close out this check-in. Just really briefly, I mean, I think it matter, philosophy systems matters because I think the philosophy, well, I guess I'm sort of a wholist in a way, at least in terms of explanation. So I think it matters because the philosophy of every little bit matters because it connects to everything else. But also this is sort of a question that can be applied at multiple scales. So it could potentially matter a lot. I guess, I mean, that doesn't really answer the question of why philosophy matters ever. But I think it's because it's part of a big theoretical framework. I'm sort of a pointy about this stuff. So these questions about realism and instrumentalism don't land all that well with me because, I mean, for example, I think the FEP, if it ends up being part of our best overall sort of account of things, then I think it's confirmed or disconfirmed by proxy whenever any observation is recorded, right? So unless you wanna say that it doesn't have any inferential connections to anything else, in which case it's really weird and I don't know why we're talking about it. Anyway, that's enough for me. Perfect, Richard and Mel to close out the section, thanks. Yeah, I think in many ways, and this is probably just maybe disciplinary specific, but the philosophy of systems and the philosophy of what we do in science, I think is incredibly important in the social sciences because so much of our concepts, so much of our theory that we deal with now is deeply indebted to Dewey and Kant and Hegel and these philosophical arguments about the relationship between mind-bodied society, what is a thing, what are the boundaries of things, and these things have been baked into our theories and many times we don't recognize that they're there and they're deeply affecting concepts and how things are operationalized and what we consider to be a truth value or not and to go in and unpack those theories, find those philosophical arguments and then interrogate them from these kind of emerging perspectives and say, was Whitehead onto something about kind of the extension of reality as an ontological category or was he not? And if he wasn't and this new principle can deal with a lot of those ideas in ways that were significantly shaped sociology, then great, we have something to kind of move on to. So I think that if philosophy is always this thing as much as I try and get rid of it, it's always coming in the back door somehow, some way. And I think just being sensitive to it, being comfortable dealing with it and interrogating it from these new perspectives, I think it's always gonna be something that's very important because it's always there in the background somewhere if you dig deep enough. Thanks, well said. And then Mel, and then we'll let Maxwell have a last thought as well. And in my mind, philosophy doesn't answer questions. It raises questions and the empirical sciences attempts to answer those questions, but you need philosophy both to, don't chew on the computer, Jesus Christ. You, sorry, you need philosophy to both raise those questions and then to tell you how those questions have been potentially answered by the empirical results. So there is no such thing as kind of science operating in a vacuum without philosophy of science. That's my opinion that there's this necessary dialectic between philosophy of science and science itself. Going on. Cool, agreed. Maxwell, and then Professor Klippinger, if you'd like. Well, so I think that there's a, there's a methodological component to this as well. I wanna echo what everyone said earlier. You know, if we're gonna study multi-scale systems, it's good to get our oncology straight. And you know, the fact that nature seems to have this multi-scale structure suggests that our theoretical accounts of nature should also reflect its multi-scale structure. But I think from a more pragmatic point of view, I like to think that a lot of the kind of single discipline topics have been covered in a very significant way, but that there isn't the same kind of, established work connecting different methodologies that operate at different scales. So I think that the getting this systems take right, not only has theoretical implications, but also has practical implications for the way that we conduct research, namely that we should aim to form multi-disciplinary groups that address questions of interest from several different angles simultaneously. So, and this is echoed in the paper that we'll be discussing today, where what we're essentially trying to do is advance on the one hand, a pluralistic ontology, a multi-scale ontology, and on the other hand, a methodological pluralism, where several different sets of methodologies can come and interact together. Nice, Professor Cliffenture, or anything else there? You're muted, but. Yes, okay. I have to apologize, because I have to leave early, but this is a great discussion, a great beginning. I just wanna throw my hat in the side. I'm sort of in a Muddy's camp and looking at explanatory power. And I think a lot of that is there. I'm so sort of ditto with what Maxwell is saying in terms of providing an integrated framework. Of course, as a philosopher, philosophy has a place in it, but I think we're there. There were so many policies embedded in historical situations that are really not, I think we're at another level of understanding that it's what becomes really important is being able to start to test these things and be able to rigorously test them. And because there are a lot of very powerful claims that are being made that if they're true, they're quite transformative, but then they need to be tested. Awesome, thank you everyone for this very exciting check-in. And I'll just add one comment on why the philosophy of systems matter. And it's examples that are brought up by Helen Longeno in her book, Studying Human Behavior, which level of analysis, which discipline we default to molecular neuroscience or environmental studies. If we default to a higher level of organization or lower level to look for our explanations philosophically, it's going to directly translate to what kinds of interventions we think are going to be possible in the system. And so if the story about, for example, aggression and sexuality, which is what Professor Longeno studies in that book, if the story is it comes from genes, well, then you're going to be less likely to undertake affordances related to the environment. If the story is about the environment, you're gonna be unlikely to take affordances related to molecular interventions. So even before we say what it actually is, and that's of course a complex question or what is the best thing for the system, I would fall on the side that it really does matter in the background, and sometimes it's just the water we swim in, how we think about these multi-level systems. So let's- Like for a three second aside, I will happily admit that I'm a staunch empiricist, somebody who worked in industry basically the moment I could. And so I, you know what, I'm just conceding a little bit that actually you're right without the philosophy of science, we don't introduce novel ideas that are worth testing in the first place, right? And that discourse is what leads to the models that allow me to try to eschew value from that quote unquote value from that. So I am backing down just a little bit from my empiricism after this discussion. So I just wanted to say that. Yeah, nice, a live conversion story. You heard it here first, but even pragmatism is defined within a use view. So it's like all models are wrong, some are useful, and then the second part of that sentence might be useful for whom or under what conception. There's no model that's just useful outside of a value set. So pragmatics isn't just like, let's just do the most useful thing. There is not just one useful thing. So after that extremely interesting check-in, let's get to what the paper is about and what the goals of the paper are. So the paper is multi-scale integration beyond internalism and externalism by Ramstead Kirchoff, Konstant and Friston from 2019. And what the paper writes their goal as is, here we focus on making, sorry, someone needs to mute. I'm gonna mute that person. Here we focus on making explicit a description of the boundaries of cognitive systems that we think follow from taking seriously the inactive embodied and extended nature of cognition. This is the idea that the boundaries of cognitive systems are nested and multiple. And that with respect to its study, cognition has no fixed or essential boundaries. And so it's all about this, taking it seriously with a multi-scale integration. So how are we going to take the multi-scale integration seriously? I'm gonna run quickly through the section headers and anyone can just raise their hand and I'll pause the section headers just to address the question. So first there's an introduction of the paper which makes sense because there's so many areas that are being tied together here. There's the first figure which is about Markov blankets and active inference. In the second section of the paper which is called a variational principle for living systems. First, they talk about this variational free energy formulation for living systems and show a figure of how the free energy principle and self-evidencing are related to living systems. They then talk about generative models and action policies because when we're talking about internal models or internal states of an organism, we're talking about generative models and action policies. They then talk a little bit more specifically in the text about Markov blankets and the boundaries of cognitive systems which recalling their goals, they're gonna take us to somewhere where these boundaries are real and they're not equally prescient as they say. However, that doesn't imply that there's one single correct level of analysis. This is related as Maxwell also brought up earlier on the formal ontology for the boundaries of cognitive systems because the modeling approach, the descriptive or action-oriented modeling approach that we take to a system is related to how we think about it. And so that's the background. If your ontology has a certain type of scheme pre-built, pre-installed, then your models are downstream of that. So when we're talking about philosophy of science, philosophy of multi-scale systems, complex systems, we're talking about what ontology will let us draw the boundaries around these systems in a way that's yes, useful, but also aware about these other aspects and the fact that we're not the only perspective on use out there. This takes us to section three on cognitive boundaries, specifically this debate, longstanding between internalism and externalism. It ends with an ism, so it's a dogmatic belief and it is just what it seems to be. It's about whether the external factors in cognition or whether the internal factors in cognition play a more relevant role or a primary role and the internal and the external are implicitly at the level of the organism. And so one of the, I think, strongest stances of this multi-scale analysis is, well, isn't there an internal and external to the group? Isn't there an internal and an external to the enacted ecological cognition like a human in a building with a computer? What does internalism and externalism mean when we're talking about niche construction, when we're talking about sociological systems? And so to explore that, first they talk about externalism in the context of radical views of cognition. So radical means from the root and so to take it from the root that externalism is the nature of cognition would be to kind of have a causal model where the outside things are more causative going in rather than the other way around. In figure three, there's a very nice abstract representation of operational closure, which is actually related to Kant and the idea of teleology and whether the organism is the fundamental unit of analysis in biology or whether other levels might be important too. And figure four shows from a auto-poetic point of view about autocatalytic closure. And so it's relating using a few different metaphors including a autocatalysis like a metabolic loop metaphor for closure as well as an informational closure that will be returned to when talking about the Markov blanket. Internalism is then given their say and it is pushing back against externalism. And if one can imagine what radical externalism looks like you can probably imagine what radical internalism looks like. And then after bringing up in section two the variational free energy principle and then in section three framing this debate about internalism and externalism in section four, we combine sections two and three which is to say that we apply the free energy principle formalism to taking it seriously integrating multi-level systems between internalism and externalism or moving beyond it. And that implies reanalyzing generative models what they are and how they're used to study cognition. A section called inactivism 2.0 which is a callback to earlier generations of inactivism. Then there's a section on nestedness or how to study cognition beyond the brain. So okay, neuroscience or behavioral neuro is about what's happening in the brain but that's nested within a body that's nested within a social group. So how does this nestedness change how we study cognition in the brain and beyond? And that's represented in figure five with this blanket of blanket ideas and in figure six with a multi-scale self-organization and active inference figure. And then the paper closes with a movement towards multidisciplinary research heuristics for cognitive science. So just as far as today's discussion goes in the next 30 minutes we're gonna try to run through the abstract and the figures and then we'll have some closing thoughts and then that will hopefully get everyone excited and interesting thinking about different types of questions which we'll be able to address in live session 5.2. In the service of the half hour we have to get through the six figures I'm just gonna put the abstract part one up here. What they're doing is presenting a multi-scale integration interpretation of the boundaries of cognitive systems and they're wanting this to be a corrective for the debate between internalist and externalist positions. The second part of the abstract they discuss how they first survey key principles of this radical extended or externalized cognitive perspective also implicitly in the context of the internalist perspective. And then they develop a multi-scale account building on the free energy principle trying to move beyond this internalism externalism debate. And then in the third part of the abstract they close up that discussion by saying that their approach does not privilege any specific given boundary of the world but it also doesn't argue that all boundaries are equally prescient. So that's gonna be something interesting to unpack. And then they close just by saying that they are drawing the boundaries of cognitive systems using this multidisciplinary research approach. So let's get to the figures and this first discussion let's just try to understand what are the figures about what is being shown in the figure just at first pass level visually what do you see but also where does it fit a little bit more broadly into the A to Z of the paper or just what did you think about the figure what did you see in the figure or what were you wondering about? So here we are in figure one would anyone like to take a first pass at describing figure one what they see or what it shows or means for the paper. All right so I'll start describing it and someone can okay Max we'll go ahead. Well so this is a depiction of a Markov blanket. And so Markov blankets are one of the kind of building blocks or ingredients that go into the free energy formalism and essentially a Markov blanket well mathematically what it does is if you have a set of random variables and you want to effectively map or individuate a random variable or a set of random variables from another set of random variables. So for instance you want to individuate a system like the brain from all the other dynamical systems with which it interacts well what you do is that you stipulatively define a third set of states between the internal states of the system that you're interested in discussing and the external states of the embedding environment. This third set of states is called a Markov blanket and it's defined by the absence of connections. So effectively the internal states are influenced by sensory states but do not influence them and external states are influenced by active states but do not influence them. So by precluding these two basically directed edges of this kind of connected graph what we're doing is essentially formalizing the kind of channel through which any interactions between the environment and the internal states of the system have to, the interactions have to go through this channel effectively. So yeah that's what the formalism does for us. It allows us to individuate a system formally in terms of the interactions that it has with its embedding environment. Great and Mel go for it. Yeah I think a lot of people publishing on the FTP in the Markov Blanket Formalism but even more so in kind of the discussions you get after a conference. People really expect of the Markov Blanket Formalism that it's picking out not token distinctions but type distinctions. So people really want the Markov Blanket Formalism to cut at natural joints. That is to find kind of the tight boundaries between natural kind, sorts of natural systems. And it really doesn't do this. It really, really doesn't do this. It really picks out individual systems. That's what it does is it finds token systems. And naturally I think a formal methodology that isn't about, it's not about finding privileged scales, privileged spatial scales or privileged time scales. It's about finding particular systems. And so I think naturally when we employ this formalism it's not going to lend us a view of privileged scales because that's not what formalism does. And I think that's kind of against, goes against what people tend to think the Markov Blanket Formalism is for. Awesome, Muddy? I was just going to say that, so I would say that most people on this call are far more intelligent than I am. And I try to come up with my own sort of baby interpretations of this, which to my knowledge nobody has been angered by before. And I think really what this tells me is if I have an agent that is interacting with an environment, the only channel by which they can actually update their beliefs in order to change their actions, say, comes through this sort of border state, these sensory states. So it's interesting in that I don't consider this, well, I mean it is a separation obviously between two objectively different realities, should we say. But it also, it speaks to the force with which, forcing the analytical link between the two through a certain set of sub-states, which is really important, right? And this is this border set on all your sensory states, we say, your sensory and active states. So that's kind of how I've always viewed this. I have two sets of realities that are trying to come together in terms of what's going on. One of them has a belief state, one of them just is, and the belief state is really trying to figure out what is, and the only way it can do that is by modifying its sensory state. So that's just something which for maybe people who aren't actively part of the community might make it a little bit more reductionist and simple to digest. But if anybody disagrees with me, please let me know. Nice, well said. Richard, and then if anyone else wants to raise their hand. Yeah, so I'm sure this is probably a naive question, but when I was listening to Maxwell talk about it and looking at this figure, do Markov Blanket's work in the sense of I hypothesize a priori that there's a thing called an environment, there's a thing called an organism, there's a thing called sensory states, and then the Markov Blanket allows me to go in and find it statistically and determine whether or not these things can be isolated empirically and therefore kind of reinforcing my idea that these things actually exist in reality, or can I just feed a Markov Blanket a whole bunch of random data and it's gonna find something, you know, kind of that garbage in garbage out model. So like when you're talking about Markov Blankets and finding these boundaries, how exactly does that work? Let's have Maxwell just directly respond to that. Very nice question. And then we'll continue on the stack with Steven and then Mel. Sure, to answer that question, it would be easier if you could change figures and go to figure, oh, let me just check on the, in the paper it's figure five, sure. So this is my favorite figure in any of our papers. And yeah, so this basically tells you what's basically going on here in the construction of Markov Blankets. So in more recent work, this has been formalized in terms of two operators, a grouping operator and a dimensionality reduction operator. And effectively what we're doing is also, to answer the question, Richard's question directly, in principle, what we've been able to do recently is develop techniques that more or less automatically, although, you know, we can discuss, there are a few design decisions that have to be made and these have been discussed in a few of our discussion channels recently. I'm sure Mel will, I'm sure you'll remind me of this, but basically what happens is, what you do is you start with a bunch of states, right? So we're working from a dynamical systems perspective and we define whatever states are of interest for us. And essentially the Markov Blankets are constructed by looking at the relative rates of change of variables with respect to each other. So mathematically what we're doing is taking the partial derivatives of each variable with respect to each other and then using effectively these rates of change to construct the Markov Blankets automatically. So you start with a bunch of states, here I'm at the bottom panel in the figure where you have essentially these nine circles, the represent states. So I mean, these can be anything whatsoever. In a recent paper that was pre-printed, this has been used to analyze fMRI data. So in that paper, each of your states is one 3D voxel. And basically what you do is you write down an adjacency matrix. So an adjacency matrix is effectively a matrix where you have all of your states by all of your states and the entries in the matrix are quantifying the effect, the influence that one state has on another. And if you take the Jacobian of this adjacency matrix, you're replacing all of these entries with the partial derivatives of one variable with respect to you others. And so having said all that, you can construct the Markov Blankets analytically by noting the absence of dependencies in this adjacency matrix. So zeros in your Jacobian mean that the partial derivative of one variable with respect to that other variable is zero, which is another way of saying that basically the two variables in question are conditionally independent in that direction. And so I mean, there are a few modeling decisions that you need to make effectively. There are two to my knowledge. One is basically gonna be deciding which states are internal states effectively. So you're gonna say, okay, if this one is an internal state, then its blanket is this and the external states relative to it are that. That arbitrariness isn't all that arbitrary in the sense that the blanket structures are gonna be essentially the same. It's just what's gonna count as an internal state might change. And then there are a few design decisions when you're implementing this. Basically you have to, when you're constructing your adjacency matrix, you have to kind of decide what counts as a cutoff for interaction. So those two things might make it less objective, but in principle, if you'll allow for these two kind of modeling decisions, you can feed these algorithms like a time series data, for example, and it'll chunk out the Markov blankets from it. Very nice answer Maxwell, thank you. So we'll go to Stephen and then Mel, then Alex Kay and Shannon. I'm gonna just get my head into it, but one of the things that I think this also is maybe not always easy to imagine is the dynamical nature of all of this, that it's talking about states. So it's not, you're never actually being one thing. You're always in this state of unfolding. And that, the Markov blanket then is that kind of statistical kind of, you could say record keeping at the different scales. Ultimately, I would say everything's at the cell level to some level because that's the kind of engine that's enabling everything else to be created. The only thing I kind of feel that is important to maybe distinguish when talking because I think there's a danger with the word systems is that there's like, I see these blankets as more like a swarming of statistics which creates the structures and the networks such as the kind of leafs or whatever. And they, and it's that kind of structures which actually are kind of in the world, so to speak. And those structures have systemic properties. They become the support mechanism for a plant, the leaf does, but the structure has come out from the swarming of the actual blankets. So I think that I'm still trying to get my head around how to pass it. Nice, we can return to that Mel and then Alex Kay. Yeah, I think Maxwell basically said it all, but the short of it is you have some time series basically and that could be real data or that could be sim data, that could be simulation, right? And you're pointing the Markov blanket formalism, you're directing it to the internal state and you're handing it your threshold for how conditionally dependent should be conditionally independent enough to count. And then that does the rest of the work for you. That will partition your state. So you give it the internal states and you give it the kind of threshold for conditional independence and that partitions the rest into external states, sensory states and active states. Cool, Alex Kay, then Shannon and Muddy. Yeah, just a really brief point. I'm just wondering if one way of thinking of this is that the Markov blanket partition insofar as it's objective together with some data, like maybe it cuts nature at particular joints, but it just doesn't cut at types of joints in a way. So I don't know, maybe we can pick that up later. Well, yeah, that's a good point. If I can just jump in here, I think Mel is right when they say that the Markov blankets are always gonna pick out individuals. If there's anything like a species level regularity that we can express or like a regularities that are specific to members of the same group, that'll be more in terms of the generative model, non-equilibrium, steady state density. So I mean, this is also discussed in the paper, but essentially what we're saying is you can imagine a joint probability distribution over all of the variables of the system and implicitly also over some fictive external states that the system is interacting with. And this entire joint density you can think of as the phenotype of the system. And it's known as a non-equilibrium steady state density or a generative model. So I think, yes, things that pick out types have more to do with this joint density that tells you about the kind of creature that I am, whereas as Mel was just pointing out, these Markov boundaries are effectively the boundaries of individuals every time. So. Cool, Shannon and then Muddy. I think I might be echoing a lot of what was said already, but on what the Markov boundary is separating, it's always going to be a decision on the part of the researcher for what correlations seem to matter for the explanation that you're looking for for the particular system's behavior. So like if there's certain long-range correlations that are constraining short-range correlations, whatever you're wanting to answer. So if I want to answer how a certain musical phrase affects the movement of the person listening to it, the particular part of the system that's interesting to me is maybe not going to be the electrical firing of a neuron at that moment because the electrical firing of a neuron at that really short-range correlation, that really short-range time scale isn't going to have a huge effect at the long-range time scale in my explanation. Even though it's important, you can't have the reaction without the brain having some electrical activity. It's still not the most informative skill of analysis. Nice, Muddy and then Mel. Awesome, so again, just bringing this back to a level of understanding that allows me to get some sort of intuition why this hierarchical Markov framework is really, really important and useful in defining dynamical systems. So one example of my background in medicine, reason why we were interested in functional neuroimaging is we wanted to understand at what scale can we intervene in certain effective connectivity architectures in order to make clinical difference. So understanding the boundary states of which pieces of this puzzle are genuinely boundary states. Genuinely, fall within some Markov blanket is really, really important. And one example of this is if I take my auditory cortex and my visual cortex at a hierarchical level of processing in here, one could say that the internal circuitry that allows me to interpret the signals from my cochlea versus my retina are not correlated with each other. However, at an integrative state above this, when I'm trying to actually make some sort of cohesive view of the world, they are both, of course, within that same Markov boundary state. So here's an example of a hierarchical separation of boundary states, which is really, really important, not only from an empirical modeling decision in terms of what level do I abstract, but also from an understanding perspective. One can say assume a boundary state and then test empirically if this is correct, I would say some sort of dynamic causal model. Simply just by changing what you think the actual Markov boundary states are, or the scopic level of that Markov boundary, you can start inferring things that are really clinically useful, for example. So the motor cortex and the cerebellum quite clearly form a Markov boundary for motor processing. But then motor cortex and S1 also form a Markov boundary in terms of action decision-making. So this tool, it's really a beautiful explanation of the world, but it's really a beautiful modeling tool as well for abstracting away hierarchical noise in terms of your model. So I just wanted to add that in. Oh, I think that's a great point. I mean, it shows the kind of meta modeling strategy that we have going on here. Cause, so in another paper that I think you've discussed on this podcast, a tale of two densities, I'm not sure that you did, you've discussed a few, anyway. In that paper, well actually, maybe we should just, I'll just seed my turn. Sorry. Yeah, no, all good, all good. Let's return to Mount and then anyone else can raise their hands. Yeah, just to quickly clarify, come straight to what Alex said. I think the term, the phrase cutting at natural joints as it's used in a philosophy of science actually refers specifically to type distinction and not token distinction. So it's actually, it's about finding natural kinds and natural kinds are definitionally type distinction and not token distinctions. Yep. Also, if I could add one note there from my field of gene expression analysis, there's often a few different ways to construct a network of relevant genes for you could pull down and see which proteins are physically touching or which transcription factors are binding to which genes. So that might be a protein-protein interaction network or a gene regulatory network. Or there are statistical networks like co-expression networks which are very related to things like Granger causality or effective connectivity. They're statistically derived and there might be two elements like two proteins or two genes or something that are correlated in their expression but never touch and those will come up on the functional connectivity matrix. Or you might have two proteins that touch and actually are very integrally important for function but they're not expressed in a correlated fashion. And so that might lead to them being missed on the co-expression graph or the co-activation graph on the brain. And so then somebody might wanna say, well, what's the real map of the territory or what's the real network? And that's where this meta modeling question returns and related to the carving of the joints. Yes, on the top of figure one we see, okay, let's carve out the organism from the environment but that's not just a really close line around the organism. That actually includes some long range statistical dependencies as well as things that might be touching but not that important for regulating organismal states. So there's multiple perspectives that are required and certainly multiple disciplines that are required when we wanna be asking how do we design interventions or systems given this causal structure. The endpoint isn't just scaffolding out this infinitely complicated reference map, it's an action-oriented framework that we wanna be talking about multi-scale models and then also knowing how to work within them. Sasha and then anyone else before we go to figure two. Great point, Daniel, about the different kinds of interactions that exist in systems and whether it's statistical or like physical correlation and I think the kind of overarching theme here is that you wouldn't know unless you measured it and you need to know what sorts of variables are important in order to even see statistically if they're going to be related. So that is where this kind of framework is really critical because we need to measure the relevant things in our system and if things are happening at scales that are too long or too short outside the bounds of our measurement then we'll never know. Nice, we'll do Maxwell and then Stephen and then move to figure two. Well, I just wanted to pick up a bit on this meta modeling thing that had been mentioned. I think it's one of the strengths of the framework to on the one hand describe like the workings of organisms using this free energy minimization stuff. So we're sort of adopting the theory itself, the free energy principle theory itself adopts an instrumentalist stance on organisms to say well what organisms are essentially doing is leveraging the statistical structure of their bodies and movement which we can interpret as a generative model as a joint density over all the system variables. So using the statistical structure of their bodies in movement to generate adaptive patterns of behavior. So at that first kind of theory level there's an instrumentalism that's going on and there's a kind of metal, there's a modeling aspect going on. We're saying that effectively organisms are sort of in a process of approximate Bayesian modeling of the causes of their sensory stimuli. Then there's a kind of meta level of analysis where we can actually as scientists write down a generative model that says well the Markov blanket is here and then score that model in terms of the variational free energy and use a variational free energy approach to kind of find where the boundary is of the variational free energy minimizing system. So I kind of take that meta aspect to be a strength of the framework. Nice, agreed. We'll go to Steven and then Mel. Yeah, and sort of following on from that as well. The piece that sometimes gets maybe lost here but is in the other papers when we look at regime of attention is that it's not just that the system cleaves like the whole, the Markov blanket gives a capacity to change how you take your attention from the stream that are coming in. Like I suddenly feel my hand, then I feel my face. And I suppose one of the things that would be an example as well of something that's at a very large scale is the threat of climate change. And I'm thinking about when I start thinking about a glacier and Antarctica starting to come away and go into the water and possibly raise in sea levels in 20 years. My physiology changes and my whole, like when I bring that into my considerations the way I take my attention and the way I tune to a lot of things around me in the world changes. So, you know, is that as my system boundary changed or I would say that it's more about my attention or integration that's changed? So that's where I think there's a question around whether the word systems could be overused sometimes and it's maybe staying with this kind of statistical dynamic is quite useful. Right, it's statistically influencing you whether you like it or not. And then your awareness and your attentional regime is something that's almost separate. Awesome, Mel? Oh, I just wanted to say goodbye because I'm supposed to be in another color right now. Thank you for participating, Mel. Awesome. Thanks for having us. This is lovely. Cool. So in the last couple of minutes I just wanted to get all the figures on the screen just so we could run through them, let people take the next week to digest them a little bit more fully come for whatever exciting parts they were interested in. So when I look at figure one, I just see on the top the simplest framing of system from surrounding. And if you think that that's a blurry boundary then you're not talking about a system. So to be able to define something is to limit is to define. So it might not be the most functional boundary but you can start there and then you ask what kinds of states or interventions connect the system to the environment just like Muddy was describing. And those are described in this model as statistical dependencies. And if we had to put English words on them they would be sensory states and active states but ultimately it's really about what the model actually says. Figure two A goes into a bit more detail in that action perception internal external two stroke engine as I think about it this feedback loop from internal and external states connected through sense and action. And then the Markov blanket is what is being pierced bi-directionally by sense and action. And as we've talked about before and in some of the work on niche construction one systems internal is another systems external. And so in some sense what is being discussed is that blanket state because there's some symmetry between the internal and the external in that both internal and external states have their own self-causing or autonomous dynamics as well as influence each other through sense and action respectively. So figure two is just showing that again in the context of a human and then also connecting it a little bit more closely to the free energy formalism of energy minus entropy. And so again another day for talking about the math there. Figure two B just talks about how those formalisms of free energy respectively complexity minus accuracy on the action bound and then divergence plus surprise on the perception bound how the math sort of sketches out another day. Figure three is an illustration of operational closure. So what they note here is that the black circles are forming part of an operationally closed network of self-organizing processes. And this is something like teleological closure. Like if you're talking about a sustainable ecosystem you don't want the sustainable ecosystem to be defined from your front door to your trash can. You're talking about operational functional closure existing at a higher level of analysis. So that means the multi-level perspective is gonna be important. But at one level of analysis or at multiple we can ask has operational closure been achieved? Have we done enough system diagramming or mapping of causal relationships or dependencies such that we actually have the system's dynamics separated out from dynamics outside of the system? If I can just jump in the motivation for including this figure is part of the argument that we're making I think that's controversial is that basically most of what the inactive approach does we can do with the active inference approach but in a way that isn't just like, well in a way that's formally backed by like a principle physics-based approach. And so we're just discussing here the notion of operational closure that inactivists use to define an auto poetic entity like a self-producing entity. And the argument that we make immediately after is that well active inference doesn't just say there is this operational closure it tells you why there is operational closure. Nice. And also to reference what we're talking about with Mel just briefly. Kant performed a teleological analysis and said that the organism is not just the limits per se of the epithelium not a physical boundary but it's actually the teleologically closed entity. And so that is really consonant with the idea that the use social insect colony is an organism not a super organism. By calling it a super organism we reify the boundary of organism being at the epithelia and people say, oh well it's like a super organism it's all these little insect organisms working together but actually because of the major evolutionary shift in you sociality in the ants for example the colony is the organism. We don't draw a little single cell organism barrier around each one of our cells and then spend all of our time running the math on the altruism between the nerve cell and the muscle cell and the sex cell and things like that. We just say, nope that's the teleological level the liver is gonna make sense in a broader end and that end is the organism. And then the question about what the privileged level of analysis should be the organismal teleological analysis says right the liver and the stomach make sense at the level of the organism but in some sense doesn't the organism only make sense in the context of its cultural niche and of its social relationship so why stop there? Where should the teleological closure end? Where should the functional closure end? And that's what demands I think a true multi-level perspective because there's gonna be questions where the functional boundary is the epithelium around the person or around the single ant nest mate but there's other times where what we're talking about is the closure around the organism as the teleological hole and that's not always the same thing as the epithelium. Yeah, Richard? Yeah, I think this general argument is really captured well in the Kramer and Darden mechanistic philosophical perspective to where a phenomenon is defined by the entities and activities that generate it or produce it and so whatever phenomenon you define whether it's the ant, it's the colony there's going to be some mechanism that generates it and that mechanism is decomposable down to whatever you want. So the colony can be decomposed down to the individual ant which can be decomposed down to the various biological structures and really when you want to move up and down you're just simply identifying, okay what are the entities and activities that generate a phenomenon that then interacts with other phenomena as the entities and activities to create that, right? And you build up and down because I think it gets really dangerous in sociology we do this a lot to where there is no individual there's only society in this teleological reasoning where it gets very hard to keep track of, okay well, Bob commits a crime not the entire society that Bob's a part of. So how can I drill down on Bob as an entity that I want to talk about but then recognize that Bob is part of a larger structure and not reducing it to the group as the entity in itself. Yes, it's 100% related to that book of Helen Longinot whether, you know, was it the arm that committed the aggression? Was it the person or was it the group? And so- It's funny, I'm sorry to kind of it's funny that you mentioned that because I mean that's kind of a symptom in schizophrenia where a lot of criminals say I didn't kill him, the bullet did and logically that makes sense in terms of that reasoning but any rational person would say, well no you pulled the trigger so therefore the bullet is responsible, you know are you responsible for the bullets pathway? Well no, I didn't kill him, the bullet did so that logic I think when pushed to its logical extreme breaks down pretty quickly but yet nonetheless I think we do tend to implicitly rely on it a lot in our analysis. Yep, and before we go to Steven it's again it's those causal stories guns don't kill people, people do well depending on cognitive diversity that statement is going to sound true or not to you and someone could say well it's not even the bullets it's actually blood loss or it's not even blood loss it's reduction in blood pressure where does the causal chain end? And so at some point we need to cut through the descriptive approach with an action oriented approach, okay are we gonna ban bullets, guns, people where do we stop? There's no single answer just from identifying the territory to action. And so these are really the big questions about causal relationships in multi-level systems. We'll go with Steven and then Muddy. Yeah I think the ability to have these mechanisms is made possible because of processes. So we think about this in art because art is seen as a process and that is creating something. So in the same way with the ants there may be a mechanism by which ants chew up and masticate pieces of wood or mud and there's a mechanism by which they build the ant colony and that ant hill has a structural integrity that allows air to flow through certain systems of movement that creates a cooling effect. However, the processes which I think the Markov blankets are working with are what make the mechanisms possible to happen like the ant is following these kind of statistical dependent processes to do what it does. And in doing so it has a mechanism just like if someone's firing the bullet there's a process the body's using to infer yet the mechanism is something we see. So I think these things are kind of entwined but they are slightly different. Agreed, muddy and then back to Maxwell. I guess this is where my original gripe with philosophy of science came from an empiricist point of view. For me, where do you snip the causal chain is something that can be empirically defined. If I have an extremely strong coupling between action A and action B such that action B doesn't occur without action A then move the Markov boundary beyond it, right? That's good. So anytime anyone says bullets don't kill people, people do or whatever, sorry, people don't kill people bullets do. I don't know what the spurious phrase is or whatever. But you can say, well, let me actually empirically look at this, what percentage of, and I'm sure there is like somebody's gonna come up with an example, oh, there was this gun that went off on itself or whatever. But statistically speaking, if I threshold that I can find basically a probability of one or very, very, or tending towards this that the gun was fired by a human being. So I think it's a bit of a, when somebody presents such a philosophical argument, I'm like, well, there is an empirical response. Science gives you these empirical tools to infer these sorts of things. So that's kind of where my original gripe about the philosophy of science is coming up. Again, I've softened it on it, but I just wanted to say, I think there's an answer to these things philosophically and it comes from something with data. Cool, Maxwell, before we finish out the figures. Cool, well, I just wanted to pick up on something that was mentioned in the chat by Richard is that the, so there isn't any, I think systematic work that relates the free energy stuff to neomechanistic philosophy a la Craver. Something I'd like to discuss with you all next week is, so as I understand it, the multi-level framework, I started off as a neomechanist, I have to say like the Craver-Bectel orientation was really something that I was pursuing and around 2013, 2014, I kind of spent a lot of time trying to see whether I was able to use the neomechanistic framework to articulate multi-level causation and my big gripe with neomechanism, at least in its most popular renditions, is that it effectively eliminates cross-scale causal effects. So there's a part of this that I buy, which is that part of my rejection of metaphysical levels comes from the neomechanist approach, but I think that, because with the story, we should discuss this maybe in another podcast, but their story is something like, well, what looks like inter-level causation is really just, it's just the same causal stream that you're re-describing at another level. One of the things I like about the FEP formalism is that while it's consonant with the broad outline of neomechanistic philosophy, it does explain the way that you have causal effects that trickle up and down. So I think that's like an outstanding issue that should potentially be discussed in a, just want to flag that for the next episode. Yeah, good call. Alex Kay, and then I'll close out the figures. All right, I was actually gonna lower my hand because Max will, so we could discuss this another time, but I guess I briefly just wanted to say, yeah, I think if you go back to Kant's definition of mechanism, which I can't recite, but basically he contrasted mechanism with systems that involve top-down sort of causation as well, like, you know, that holes can affect parts. I would guess you could still get some kind of mechanistic description at each scale, like approximately or something, but anyway, it's interesting topic. Oh yeah, yeah, just to be clear, my view is that the FEP is broadly consonant with and is in fact an exemplar of the neomechanistic research strategy as outlined by Bechtel and Kraber and so on. But I think it does impose a different metaphysics of levels than the one that's been attempted. At least in a few papers that I saw by Bechtel and company, like the one called looking up, down and around, I think, where they essentially say that, well, there's no such thing as cross-scale influences. There are really just big things interacting with small things. And that's where I think I strongly disagree for reasons that hopefully will impact next weekend that are related to how these multi-scale systems are actually constructed. And I mean, essentially you've got the same base system, the tacting, so it's like, I don't know, it's a lady-min and Ross style. There are only basically quantum fluctuations or whatever, but what you're doing is whenever you define a new scale of interactions, you're adding a slower and spatially larger set of constraints on the same fundamental set of things. So there is something like cross-scale talk in the sense that what it is to be a multi-scale system is essentially to have components that are constrained by successively higher level constraints on the dynamics. But I don't think ultimately this conflicts with the philosophical perspective of neomechanism, which ultimately at the end of the day just says, well, there are parts and processes and a mechanism is a coordination and an organization of parts and processes that produces some kind of interesting output that we can study, so. Cool, yep, these are Alejandra, go ahead. Yeah, yeah, thank you. And just following up this idea you're saying, Max, well, I was thinking is there any differences in how these hierarchies in terms of blankets work? I was wondering, at the cellular level, maybe belittling is kind of risky. So the cell has to behave like having this generative model, but it's like it's phenotype, but it's not learning. I don't know, maybe I don't have very clear what I'm thinking, but if you go up to the scale thinking like in the brain level, the blanket, now belief of dating, it's crucial for this bodily interactions and effective bodily interactions and this coupling with the environment. So my question is, is there any differences in terms of how these micro blankets and blankets of blankets work? And it's also a hierarchy and how this hierarchy belief of dating works. Let's return to this great question when we have a lot more time next week, which is about the similarities and differences with different systems. It's really something we should consider. Because- I was gonna say, this is like the question that we've got to address in this paper in large part. I mean, a quick answer to your question is that formally speaking, there's a kind of self-similar pattern that is occurring across all these different scales, right? So you have a non-equilibrium steady state density and a generative, slash generative model and that is kind of directing policy selection in a way that maintains the blanket. But that's a very, very abstract or information theoretic description. And in effect, basically the kind of data that's involved at every scale, the mechanisms involved to make sure that your data, the data that you're sensing fits your preferred data distributions. All these things are gonna be radically different from one scale to the next. I mean, yeah, clearly the way that we update beliefs about ourselves by having a conversation, while it has a formal, while it may formally be identical on a kind of process or information flow level to like what our cells are doing, like clearly this is different, right? Like the data streams with which we're dealing are different, the ways that they're being managed are different. So yeah, I mean, there is a kind of fractal self-similar structure, but it's at the level of the information flows, not so much what makes up. So in mechanistic terms, the parts are different, but the processes might be different. Yep, and it will be awesome to go into more detail because there's clearly differences between cells that embody their beliefs through morphology that might be slower time changing and brains that can rapidly emulate or at least simulate various phenotypes being neural states. And in a previous paper we talked about whether those neural states are themselves phenotypes, just to quickly run through these last figures. Figure four is an example of auto-catalytic closure. So this could be like a metabolic network or self-replicating nucleotide sequences or programs that replicate themselves with or without mutation. And the point here is really to say that that kind of causal dependency structure that we saw in figure three, it might not just be bumper cars or ants bumping into each other. It could be something like different types of information that are connected to each other. And so as we start to think about different spatial temporal scales and all these different mechanisms that exist, we wanna find those invariances at a structural and a dynamical levels that help us study this diversity of systems using a common formal framework. Figure five is something we already looked a little bit at, but it is this fractal idea zooming in on the fractal broccoli and basically the left side has sensory and action states surrounding three gray nodes, but that's itself zoomed in from a swarm of similar agents. And so in a tissue, maybe it's very patterned, like in the skin cells, the boundaries between the cells are very stable so they can have adhesive molecules that make it a functional skin sheet and they're still exchanging information in that grid topology, but also maybe the swarm on the right side consists of moving subunits. So like Steven was talking about with a swarm and at that point, especially, it is important to have a statistically driven understanding of causal relationships rather than one that's just around drawing a dashed line around the system as we visually see it. And then just lastly, figure six, which I think will be perfect to save for later, especially because it in some ways does connect this action inference generalization to specifically morphogenesis, which is an example of a phenotype decision-making that people might not always expect to be put in the same category along with what it is our brains or CPUs are doing, but I think that's a perfect jumping off point for next week when we, yeah, yeah, nice. And going through this is going to hopefully address Sandra's question, I think. So yeah, this is a good jumping off point, I think. Perfect, and we have a few questions just prepared for next week, but we're not going to go to them now. We will talk about that next time. That concludes the main section and we'll also just have anyone can raise their hand and give a final point, but I just wanted to say that we'll provide a follow-up form to our live participants. I'll put it in this jitsie chat as soon as someone starts to give a final thought. Any feedback, suggestions or questions from our listeners, participants are always welcome. Of course, everyone's engagement is welcomed and people can stay in communication with us. So if anyone has their hand raised, I'll be happy to hear their first thoughts while I post the link. Nobody's hand is raised, but thanks again for the continued engagement with our work. It means a lot and I think this is extremely useful. Like both I think for us as a little community working through these things and for the broader audience of the free energy active inference stuff. I'm very grateful. Thank you very much. Awesome, Muddy raised your hand and if anyone else wants to give any final thoughts what they reflected on from today's discussion, what they'd be excited to hear about in next week discussion. I was just gonna say thanks to you. This is really, really great. I think first of all, just on a personal level, it's great to get back involved with people who are much clever than I am and be surrounded by this new knowledge. But secondly, I can really start to sense the snowball effect and the active inference community. It's only a matter of time before it really leads over into industry. It's a really, really exciting time to get involved in the space. Thank you all, that's really it. Cool, Richard and anyone else freely raise your hand. Yeah, I think this is really interesting. I think there's so many different points of connection and interaction with so many different disciplines. And I think one of the difficulties is the barrier to entry is just so high in terms of the terminology, the mathematics, to try and rap a heater, head around these things. But once you begin to understand it conceptually, I think people can start to go, oh, okay, I know what this is and I can talk about this. And so I think one thing to maybe discuss moving forward is are there ways to kind of get almost sandbox working groups where people can kind of start really empirically playing around with ideas and data sets that kind of emerge naturally out of these conversations because cold calling somebody and saying, hey, I've got this idea. Can you explain to me what a Markov link it is and what's worked together? People are like, nah, son, I got things to do, I'm sorry. I think there's a way that we can kind of merge from these really great productive working group or discussions into maybe something where people start producing actual collaborations across disciplines. I think would be a really fantastic thing. Awesome. I just want to reiterate also that there's this reading group that's ongoing on the Carl Friston's monograph. So for those of you who are interested in actually surmounting the barrier to entry, as it were, we've recorded the chats and everything. So if you do want to join, just drop me a DM. And yeah, we've grown to, this was originally going to be like me and eight buddies just reading through the monograph and we've grown to a community of almost 100 scholars. Yeah, following this remotely with the recordings available and we're going to turn these recordings heavily edited and everything into pedagogical material, kind of like video based a companion to a free energy principle for a particular physics. So yeah, if anyone is interested, the group is kind of semi loosely open especially to people who are doing active research in the tradition. So yeah, drop me a DM and I'll hook you up. Cool, Steven, and then anyone else if your last thought comes to mind. Yeah, well, just to reiterate, yeah, it's really useful. I like to see everyone nibbling at this from different directions. It shows a lot the questions everyone's asking. It's like, because otherwise you're thinking, it's just me that's thinking this and sometimes it is and then sometimes it's not. So it's kind of good to know which is which because sometimes it's not clear, it's just a figment of my imagination. And so the only other thing I thought might be cool to bring is that new paper that came out on reciprocal active influence that's just come out might be interesting at some point to think about because that might open up another level of development from this paper. Yeah, very interesting paper by Matt Sims at Edinburgh who I'm very excited about this stuff. So Matt Sims developed an account of symbiosis that draws on the multi-scale integration framework that we've developed where he talks about bi-directional multi-scale integration. So not just integration of the parts of one biological organism, but the integration of effectively a symbiotic union of organisms with different phenotypes. The paper uses the, I think it's the Bob-Tales squid that gets colonized by bacteria. And so yeah, it's a very, very interesting paper and I'd highly recommend it. Cool, cool, very interesting. Sasha? Yeah, I just wanted to reiterate what Steven just mentioned. It's great to have this ongoing conversation and to hear how different people phrase the questions that I think we're all thinking. So I'm really looking forward to next week's discussion and kind of seated with the question that Alejandra stated earlier. Perfect, well, thank you all for participating. The Google form for you to provide any feedback for the live participants is in the chat. For people who are watching this live or in replay, please provide us your feedback so we can make this better on-ramp for active inference, the community, the ideas, and ultimately the projects too. So it's just great stuff. Another one in the books, 5.1 for active and I'm gonna terminate the live stream. Thanks so much for listening. Thanks.