 Right. Hello and welcome. This is Active Inference Livestream 63.1. It's November 7th, 2023. We're here with Mikhail Bikarski and we'll be hearing a presentation followed by a discussion on incorporating variational free energy models into mechanisms, the case of predictive processing under the free energy principle. So looking forward to this lecture and discussion. Thank you for joining and off to you. Thank you very much, Daniel. Thank you for invitation. This is a great corner for me and very nice. I'm a little stressed. I hope my presentation will be interesting for you and give a chance to take some new ideas concerning the Bayesian models, especially predictive processing under the free energy principle. So my talk is based on my recent papers, which I publish after many, many months of revisions in CENTES, and I will sketch the view which I try to develop here, which I think has some interesting implication also for the discussion in the field of Bayesian modeling, like in the wider context of philosophy of neuroscience or philosophy of cognitive science. So I will try to justify the three hypotheses, one general and two more specific concerning, surely the predictive processing and free energy principle. So the first one idea which I found in a recent paper of some philosophers of mechanistic explanations is the idea that there are phenomena, I think that there are neurocognitive phenomena, which should be explained not only in the terms of the composing the mechanism which is related to this phenomena, but also should incorporate the constraints from the environmental constraints and the flows of free energy. I will defend the view according to which if we want to explain those mechanisms, those phenomena, we have to take an account of flows of free energy. And the notion of free energy of course in the work of those mechanists is related to the thermodynamic free energy, but I will try to show that the free energy we can understand also in the terms of informational or variational free energy. This is the first hypothesis, the second hypothesis is this is not my idea because we found in many papers of Jacob Hoey of or Paweł Gładziejski, the idea that predictive processing provides a sketch of mechanism. In other words, predictive processing is a computational framework which has mechanistic or in which can offer a mechanistic explanation of its target phenomena. And I will try to connect the two in my hypothesis and shows that mechanistic predictive processing takes informational constraints from the free energy principle. In other words, we can integrate free energy principle as normative theory, normative principle with process theory predictive processing using the mechanistic picture, mechanistic idea of explanations. So let's start from the first hypothesis. I think that there is a general and common agree in the works of some authors from cognitive science from philosophy of neuroscience. The view which is schematically called a new mechanical philosophy according to which the explanation of phenomena is a matter of the composing this phenomena into underlying mechanism. So when we want to explain phenomena, for example, this is the view to defend the the Carl Craver or William Bechtel and the others, when we want to explain phenomena, not only cognitive but also biological social phenomena, we have to decompose the mechanism which is responsible for the realization of this phenomenon. And this mechanism can be treated as a mechanism which casual product the phenomena or a mechanism which is responsible in the terms of constitutive responsibility for the realization of the phenomena. And those authors and those traditions which Craver in his book explaining the brain calls the system traditions, claims that explanation is a matter of finding the systems, the mechanism which should be described, should be identified by scientists and should be decomposed into parts, into components. We should recognize the relation between those components and we should recognize and describe the organization which rules of the relations between the components. And it means that we can think about the mechanism in terms of hierarchical organization because if we look on this figure from the book of the Craver, phenomena are based hierarchically on the underlying structure of hierarchical structure of a mechanism that are lower and higher level mechanism. And according to our scientific aims, according to our strategy, we can looking for the more deeper or more deeper mechanism concerning, for example, how the level of biological organization to the level of the physiological organization and so on. And one of the most important authors in the field of mechanistic explanation, the field of new mechanism, William Bechtel, in recent latest paper, he little change the standard view on the composition. He claims that heuristic of the composition or normative strategy of the composition and mechanical explanation should be rethink, we should rethink this strategy. Why? Because there are high level cognitive mechanisms, which cannot be described solely in the terms of hierarchical organization or in the terms of casual relations, because there are some kinds of mechanism, which are as Bechtel and Colex claims, which are active control mechanism. There are mechanisms which are responsible for production, the phenomena. What kind of phenomena, the active phenomena, active structures like bodily movements responsible for body movements and physiological process and Bechtel claims, sorry, that those mechanisms, of course, they play a central casual role in the hierarchical organization, but they also a components of complex network of heterarchical web of control systems. And they play its casual role. They casual role because they perform a given function in this heterarchical web. Why it is possible? Because in this sense, those phenomena are mutually constrained by the other elements, we can call their constraints, and they are active or they derive their casual efficiency because they are constrained systems. And in opinion of Bechtel and other authors who co-work with him. It is important to rethink our old image of the composition because in standard view on the decomposition, we don't take an account on two elements on constraints, which make given a mechanism active in the sense of the being part of heterarchical web. The second point is the energy because biological mechanisms, cognitive mechanisms are always related to the flows of the free energy which they transform, which they process. And why this is important? Because according to Bechtel, he claims, what is characteristic for biological mechanism in the context, in the contrast to the other mechanisms, is that the biological mechanisms are dissipative structures. It means that they occur and maintained in the context in which free energy is being dissipated. And if we don't incorporate the element of transforming or processing the flows of free energy, in the fact that we don't really explain why given mechanisms, why given systems have behaviors. As Bechtel and Bechtel claims, completely unconstrained system will have no behaviors. And I will think that if we want to explain neurocognitive mechanism, we have to adopt the view which I describe in my paper as the constraint-based mechanism approach. And according to Bechtel, it's a heuristic, mechanistic explanation of at least some cognitive phenomena should be based or should incorporate dimensions of constraints, element of the environment, which makes mechanism active, and the flows of free energy. And this is the first piece of my idea. And second is that I think that we can try to demonstrate this idea in one example from contemporary cognitive science, and I will try to adopt this idea into the frame of predictive processing. So now I will sketch a short view about predictive processing framework and why this is important for me to try to test this framework into the context of the constraint-based mechanism approach. So predictive coders develop the Kantian, the Helmholtzian idea that our cognitive systems, our brains, don't have a direct access to the causes of information we receive by the input. And they claim that the central, the main function of the brain on the surgical nervous system is to minimize uncertainty, which is described in the terms of minimizing long-term minimization of average prediction error. How is this possible? How brain can do this? It is possible because the brain inhabits or embedded the virtual informational structure, which predictive coders by Asians describe as a generative model. And the generative model is hierarchical and multi-level structures, which transform, process the information coming from the input. Why? Because in this way, the brain can minimize the so-called prediction error, the discrepancies, the differences between the state of the model, state of the organism, and the information coming from the environment. Because successful minimization of prediction error has many adaptive benefits. So this is a general picture. Now we should look more closer to this image and try to develop a more formal technical language. So we can use the idea which develops a lot to think about the relation between the events in the world in the term of relation between cause and effect. But what we receive is the observation we receive by the sensory input observations. And if the idea developed by predictive coders coming from Helmholtz is that we don't have an access to the straight true states as understood as a causes of the observation. So we have to infer the true states to act successfully in the environment. And how it is possible to infer to guess the true states responsible for the observation we receive. So there is a development of the Helmholtz idea of unconscious inference. And Bayerzian claims that this idea can be translated into the thinking about the statistical inference. What kind of statistical inference? As we know, this is the Bayerzian inference. But what does the generative model minimize prediction error or infer the true states using the Bayerzian inference? The Bayerzian inference rises a small problem because Bayerzian inference, exact Bayerzian inference, is intractable hard. You cannot compute exact Bayerzian inference. And the idea which developed, for example, Carfriston is that our models, our brains using models, generative model, don't compute exact Bayerzian inference, but it uses the simpler version of Bayerzian inference. They approximate, they use the approximation of Bayerzian inference. So in other words, if the story is correct, then continuous minimization of prediction error can be described or understood or explained in the terms of approximation of Bayerzian inference. And this is, this approximation is possible because model minimizing prediction error simultaneously minimize a certain quantity, which Friston calls variational life energy, the quantity which is always greater than or equal to suprisor. And using this quantity, model iteratively update the internal parameters under the internal states. And in this, in that way, can infer the most probable true states, which is responsible for the observation which model refers. So according to predictive processing, as I read it correctly, what is characteristic for the brain is that the brain main function is to processing information in the terms of Bayerzian approximation. And I try to unpack this idea and read this and try to read this in the context of a new mechanical philosophy. So this is the first step. The second one is to justify that predictive processing can be connected with mechanical philosophy. And in many papers of Wawegwajiewski, Hortnesk, Shava, Howie, we see good arguments according to which predictive processing as a computational architecture provides the scientist's so-called sketch of mechanisms. What is sketch of mechanism? This is the idea developed by Piccinini and Kaplan and they claim that sketch of the mechanism is the incomplete representation of mechanism. It means this is the some kind of representation of mechanism when the scientist omitted some components, some parts of the mechanism. And good sketch of the mechanism should lead us to the schema of mechanism, which is the complete representation of constitutive element of given mechanism. So if predictive processing gives us a chance to write a sketch of the mechanism, and if predictive processing is about the neural function of the brain, then if the view, which I developed after the ideas of Bechtel and Colex, if this view of the constraint-based mechanism approach is correct, then I will claim that if predictive processing, if predictive coders want to build a mechanistic explanation of cognitive phenomena, they should include into the explanation flows of free energy. Because without this flows, without the constraint of information of, in general, in free energy, we cannot explain how it is possible that mechanism operates and is active as long as this energy is available. And the thesis which I defend in this paper incentives is that predictive processing models should include informational constraints if they are to be mechanistic. Of course, if we think about predictive processing as the way to explain the function of the brain. And of course, there is one natural questions because many authors will claim that the good theoretical normative principle for building models in predictive processing offer the free energy principle. So the question arises, does the free energy principle can offer these constraints for predictive architecture? And they will be claimed that yet, yes, the FEP offers these constraints, but of course, we should try to justify this thesis more specifically. But we find the one big problem. When we look into the 2018 paper of William Bechtel, when he introduced the idea of including the free energy flows into explanations, he directly write that the notion of free energy, which he developed in his papers, in his view, is distinct from the free energy principle or free energy articulated by Carl Friston and the others. In other words, he said that, okay, we should include into explanations flows of the free energy, but this is the thermodynamic free energy. We cannot use the other kind of the energy because this is distinct notions. In the same year, in 2018, Friston published a draft or a preprint of his book of free energy principle for a particular freezing and he said something what contradicts with the view of the Bechtel. He said that he said that free energy in the thermo dynamical free energy is consistent with the variational free energy. It means that Friston claims something what contradicts the thesis of the Bechtel. And I think that it is possible to justify the thesis of Friston, but possibly I try to do this in a different way than Friston does this. For me, as a person without good formal preparation to reading the paper of Carl Friston, it was hard to find a good way reading of his paper on his co-operators. But I think that we can try to reconstruct the idea which supporters of the free energy principle develop and this ideas can be commonly described as something what I call the Bayesian mechanics argument. The first site I wrote here is coming from the paper of Max Ramstad and he claims that at the core of Bayesian mechanics is the variational free energy principle. So that Friston in his latest paper proposed the idea of the looping of the Bayesian mechanics. The Bayesian mechanics which can investigate the scribe and offer the article background for explanation what is the principle, what is the variational free energy. Why this idea is important for me? Because Friston will claim that every kind of mechanism each mechanism can be traded as an expression of some kind of mechanics. We've got classical mechanics, we've got quantum mechanics, thermodynamic, statistical also Bayesian mechanics and every kind of mechanics has its own reified constructs. Those constructs can be understood as thermodynamic energy, temperature or variational free energy and what is important is that those constructs, those reified constructs, those notions are justified and can be used only in the realm of the given type of mechanics. It means that it is possible that we've got some reified constructs which we can use on the base in the realm of the one kind of the mechanics but we cannot use this in the other type of mechanics. For example, when we talk about the denotion of the temperature, it will be used in a different way in the realm of quantum and in the realm of statistical mechanics. It means that if we take the construct like variational free energy on thermodynamic free energy from the point of view of this argument, it means that all of these constructs are relativized or related to the description and method of measurement which we use in the specific type of mechanics. What does it mean? It means that there is no ontological, we don't have a good ontological argument to claim something like that. For example, statistical mechanics is more primary than, for example, classical or variational mechanics. Why? Because every kind of mechanism, each type of mechanics can be traded as a complementary description of the behavior of a dynamic system. It means that the most primary, the most basic structure we try to explain describe in terms of mechanics, dynamical systems or the dynamical systems. According to Freistone, it means that Bayesian mechanics can be traded as all others kind of mechanics, which we find in physics. But what is specific for this mechanic, as I said, is the assumption of variational free energy which is strictly related to the assumption of Markov blankets. Freistone will claim that these additional constraints, the constraints of Markov blankets, makes possible to speak of states of something as relative as something different of something else. And why this is important to make a possibility to think, to talk about, for example, internal and external states? Because this is the possibility of talking about the different states directly applicable and directly important to the living organism and, for example, narrow structures. Freistone in his book claimed that every type of mechanics ignore assumption of a Markov blanket and the assumption of variational free energy and in some implicit sense will claim that the state outside blanket and the state outside blanket, this difference between those kinds of states can be ignored. Why? Because if this, why I'm thinking this way, every kind of mechanics give us a chance, give us a possibility to talk about the systems we tried to describe in terms of heatbed or thermal reservoir. It means that every kind of mechanics, except quantum mechanics, give us a possibility to talk about the states as in equilibrium steady state. According to this argument, we should say that variational free energy can be applied and justified only in the realm of Bayesian mechanics. And this application helped us to describe, explain why some things are autonomous or active, why some things can be as spectral sense dissipative structures. But it means that variational free energies give us a chance to describe some kind of phenomena, some kind of system as autonomous or active, but and simultaneously thermodynamical free energy can be implied in the realm of the object which we try, which we want to explain in the terms of statistics as a statistical ensemble. And this is, this is why Friston will claim that both variational free energy and thermodynamical free energy are consistent because they are two consequences or expression we can say, the same more elemental thing, the thing which can more elemental, more basic mechanistic or maybe quantum nature. And it means that thermodynamical free energy and variational free energy are the notion or constructs which we can trade as the as a two side of the one coin. And using them, it's dependent of our method of measurements of our scientific interest from the our strategy to investigate phenomena, but they are complementary in the sense that they offer the scientists that the possibility of describe dynamical systems from the two other sides, but those sites are complementary. And how we should under understand and interpretate the Bayesian mechanics argument. Because I think this argument should be interpreted in a more philosophical, more specific way. If the thermodynamical free energy and variational free energy are two sides or are related to us to aspect of more primal dynamics, there is still open the questions about what we talk when we talk about systems which minimize variational free energy and when we talk about the flow of the thermodynamical free energy. And I think that ultra realistic or super realistic interpretation, literal interpretation of this argument is not correct, because it seems that we should interpretate construct like like thermodynamical and variational free energy as the useful useful function which are which uses which justification are related to our scientific models and so on. In other ways, there will be the problem to defend the thesis of free stones that the Bayesian and stochastic mechanics are equivalent formulation of the same thing. I think the same thing about which with free stone talk is the dynamical systems. And we try to find a good tools to describe the systems, but it means that we don't have a good tools which represent or map the real structure of the systems. And it means that the notion we use like rational free energy and thermodynamical free energy should be understood as the constructs which are related and ultimately reduced to the method of measurement descriptions the method we use. So it means that when we talk about the consistency or complementarity of variational free energy and thermodynamical free energy, we don't have any we don't make any ontological commitments, any ontological assumption about the representational architecture propers of the target systems. It means that the models are which are based on the free energy principle address the cultural structure of the world in the sense that they are epistemically useful. We can use them in the practice of model building, but they are not epistemically useful in the sense that we describe the real properties which we can literally describe as generative models as variational free energy and so on and so on. And I think when we talk about the free energy principle the instrumental and realistic interpretation is correct and the free energy principle doesn't imply any ontological solution about the target phenomena. But from the point of view of my argument there is one problem because this instrumental interpretation of Fristonian argument doesn't have any mechanistic, specific mechanistic implication. We don't anything talk about mechanistic character of the phenomena. Why this is important because when we come back to the field of philosophy of new mechanical, the field of new mechanical philosophy, we all see that I think there is in very general there is a common view that we should think about mechanism in a realistic manner. What does it mean? The view which I try to describe here is so-called mechanistic realism according to which we should think about the structures, the entities in the world as the structure which are in some sense richer than their aggregation of causes. Why this is important? Because mechanism which we try to identify the compose should be produced by some kind of structures and these structures should be more richer than a simple aggregation of the causes. This is the first part of this argument and the second connected to the idea I try to develop is that there are some of those structures whose at first this organization cannot be reduced to an aggregation of causes and second this organization should be explained in terms of mechanism which are constrained by a flows of information flows of thermodynamic energy and describe maybe in terms of minimizing rational energy. In other words I will reclaim that if we adopt the mechanistic realism we should try to say that there are such phenomena such structures the explanation of which should take into account the energetic constraint which we can try to describe in terms of rational free energy. If we try to adopt the image of coming from the mechanistic realism we will defend the view which I describe as moderate realism by Asian modeling or moderate realism processing and the free energy principle. In this moderate realistic interpretations we can say that system minimizes prediction error or minimizes variational free energy because it implements in some way some casual mechanism that can be described approximately in terms of minimizing variational free energy or simultaneously maximizing mutual information between internal states and sensory states and it means we can identify describe and decompose some kind of mechanism which we should trade as a system of mutual constraints that restrict the flow of information the flows also the thermodynamic free energy to perform work in such a way that those mechanism minimize or should minimize discrepancy between internal state between the parameter of the model between the prediction of the model and the information coming from the environment in which given systems act and live. Why? Because without taking those constraints flows of information we cannot describe and explain why those kind of mechanism or why this kind of organism are at non-equilibrium study states. So and I think that we can find in the field of thermodynamics of information and also in the work of graph system argument which I called argument from narrow computation a good argument for the moderate realistic interpretation of the principle and according to this argument we can say that there is always trade-off between neural information processing and thermodynamic energy post it means that every time when biological sensor detect a change in the environment every time when biological organ process the informations this processing confirmation or detection a change in the environment is proportional to the amount of information is proportional to the some minimum energy in term of thermodynamics so every time when a system transfer process information every time there is a minimum cost of thermodynamic energy and this argument from neural computation could be this related to the formal investigations from coming from the Friston monograph about the Erzinskii equality piece of this book about the Erzinskii equality is really hard but one and a half years ago I was stuck with Carl Friston about this argument about the Erzinskii equality and I think we can try to reformulate in a more non-formal way and according to the Erzinskii equality if we interpret this in the realm of the free energy principle this equality allows us to connect variational free energy with thermodynamic free energy in this sense that when we want to describe a moving one system from one state to another and this moving in the system we can understood as creating a destroying the information there is a certain amount of thermodynamic work cost and thermodynamic action which should be entailed by this moving in other words when we look into the generator model in our brains we will say that every time when the model update belief in the in the sense that every time when model change information in the internal state of the model every time there is a thermo dynamical work cost so the belief updating in the Bayerian generative model is strictly related to the thermodynamic cost of this belief updating and I think that moderate realistic interpretation should explain us how it's possible because because when we talk about the belief updating generative model when we talk about the minimization of error literally we don't map the formal structures we use under the target phenomena I don't claim that there are real generative model in our brains I don't think that our brains optimize based and so on but I will claim that there are some structures in our head there are some structures in the one that are richer than simple aggregation of cost and those structures implement some casual mechanism that scientists can approximately describes in terms of models which minimize rational energy or prediction error and I agree with the thesis of Kirchhoff and Colex that our models fit the data without literal mapping yes our model is are true because they are approximation of the data and I think that we can think about the predictive processing model incorporated or related or integrated sorry with the free energy principle as some kind of approximation of the data which we want to investigate a specific type of the data like our neural organs and so on and I think that we should distinguish and remember about three distinct elements the fact the free energy principle as former principle the former principle which doesn't assume any ontological assumption and doesn't imply any ontological commitments about the target phenomena the predictive processing framework as mechanistic computational models which are granted from the one side in this former principle which the former when the former principle offers the variational methods variational notions and from the other side on the heuristic of mechanistic explanations and the first the biological systems the target systems that predictive processing and by Asians employed to model which are based which are independent of the implication or on the language of the free energy principle and also which are independent from the language and ontological implication of process theory like predictive processing so I'm going to conclusion so I will claim that the reference to the informational constraints allows the scientists to explain the neural organs the brains not only as a kind of physical material physical mechanisms but also as systems which have homoesthetic futures homoesthetic nature so and I will be claims that it means that a full satisfactory explanation of how the brain works not only how the brain can be traded in terms of statistical or classical mechanics but also with the terms of thinking about the brain as active as and the part of living body a full explanation of brain as part of the living body requires taking into account informational constraints which I believe can be characterized on the slide in terms of Bayesian operational free energy minimization and the last one conclusion is the observation that from the point of the view of the argument I try to develop is that the free energy principle is also normative in the sense that it sets a norm that should be met by mechanistically non-trivial process theory or predictive processing models assuming the correctness of the constraint based mechanism approach okay thank you very much that that's all what I have thank you great presentation I'll make a first comment meanwhile anyone who's watching live is welcome to ask a question okay well I would like to applaud this great talk you connected the scholarship and the framings that are used in the philosophy of mechanistic accounts with free energy principle active inference predictive processing and I think you articulated several lines of arguments that were in the literature including where there was change or even partial incoherence or just partial ambiguity certainly ambiguity related to justifications or understandings of the relationship between thermodynamic free energy and mechanics and Bayesian free energy and mechanics and that is a very wide space with everyone saying things like it's a purely formal resemblance and it's just a convenient computational heuristic used in science and physics and engineering and there's just no need to worry about more than that ranging to arguments that that they are co-extent or that they have different kinds of statuses so I very much appreciate in your paper and in this presentation that according to the way that different discussions about mechanism have been carried out in the biological sciences using that setup to clarify some of these discussions in this field that are all great to approach with kind of a first principles view and it's helpful to see how it connects with the arguments of like Craver and Bechtel Thanks I agree that I try to develop some arguments which exist in the current literature but for example when I read the paper of Bechtel I think that this idea doesn't be directly explicitly articulated and I think that Bechtel should be doesn't agree with my argument my interpretation of his view because of the variational free energy but I still believe that this is very very basic intuition but this intuition says that if we if we think about the part of the human organism if we think about the brains what is more specific for the brains is that the brain process information not brain process the thermodynamic energy not brain a process or produce some kind of hormones but the brain process the information this information is very important for for a living organism and I think that this intuition should be tasted in the field of in the field of science or maybe in the field of philosophy of neuroscience and I'm not so sure that this is possible to taste it but I think that this is important to the description or explanation of those mechanisms that those mechanisms as the are the mechanism of transform processing information and possibly possibly maybe it's enough to say that okay those mechanisms the formation with the brain with the brain process it's not important to explanation of how it is possible but I think that predictive colors the the the supporters of the free energy principle sense something important and they claims that what is specific for the brain is not only that the process information but also is the process in some specific way using something what we can describe in terms of Bayesian proclamation or more wider as a statistical inference and I think that and I think if predictive coders predictive processing is a fruitful strategy of explanation in the field of neuroscience if the model builds by predictive coders are scientifically useful then we should take an we should use this heuristic coming from the free energy principle that take it seriously because I don't want to I try to say that talking about the brightness which uses which uses which uses the generative model which approximates Bayesian inference this is not the way of talking we are speaking I think that without only describe but but this the word we use the notion we use as I said in some way corresponds to the some kind of structures casual structures which we are which we have in our head and maybe we investigate we invite more specific better words better concepts better tools to describe those structures but I believe that for now predictive processing under the free energy principle give us a good chance to explain why our brains are not only the matter but also the part the important part of the living bodies and what is the difference between the power the brain as active control mechanism and why this brain shouldn't be treated as a heat bath or thermal reservoir and so on and I think that and I think this is the way this is the reason why Bechtel tried to reformulate this this classical image of the composition because he observes that some mechanism cannot be described or treated as non-living non-active and if they should be treated as active they should be treated as active and living then I think we have to we have to try to incorporate into our understanding into explanation the fact that the information which the brains those mechanism process is really really important to understand why and how they act how they are the part of dissipative structures like our organ. Thank you I'll ask a question next from the live chat John Athos wrote what are you defining as information when you say the brain is processing information? Good questions. I think when I talk about information I think about some kind of semantic information. Of course if we try to use in practice this idea which I try to develop the question about information should be more specific more information should be more specified but I find the argument which developed Rosa Kauai as I remember when she claims is that when we talk about the information which is processed by single neuron of course this is the some kind of syntactic information there is no any semantic but which when we want to try to understand how it is possible that the brain as full organ cortex process information this is always strictly related to the in some kind of semantic information and I think we can think about this semantic information and try to but this is I don't know for now how it is possible in the realm of my argument but we think to try about this what we traditionally describe as semantic information in terms of variational information because I think that if brains minimize prediction error if brains approximate Bayesian inference it means that brain operates uses the information in semantic sense. In my paper I wrote about the mutual information between the internal and sensory states but I think I see this as a lack of my view because I don't emphasize what kind of information is precisely I mean but I think that if we take this view as some kind of proposition of heuristic of building the models the empirical investigations lead us to them to the possibility of describing this information more specific more specific terms. Awesome very interesting so a few lines that I'd love to ask questions about so in your paper and presentation you center the variational free energy as kind of the analytical construct doing the work in the mechanistic account and I wondered what leads to one analytical constructs doing that kind of work for example why not center surprise and see variational free energy as a computational heuristic or just an approximator so what would be the relevance in a in a mechanistic account that centers the concept of surprise or the concept of prediction error versus the concept of variational free energy which are related concepts in that an organism minimizing one does minimize across all but they are different in terms of their formal basis and what we could say about them. Okay I think that I think that if the answer on your question can be can have two parts the first one is related to the view which I defend as moderate realism and I will claim that notion we use the concept we use are related our core response to the structures which we try to develop and describe using these notions but it not the sense that we literally map the content of these notions onto those those target phenomena so probably variational free energy for now looks a good fit to those those phenomena from my from my point of view but it doesn't mean that we don't can use the other notions but you ask why why we don't talk about surprise or predictive errors I think that we don't explain in a in a satisfactory way this idea when we talk about the surprise maybe our prediction error because we don't understand why why brain approximates by general inference if we don't introduce the notion of the variational free energy or maybe it's equivalent to the upper bound of surprise but if we don't introduce the solution then we don't find the computational formulation or computational foundation to description of the brain working in terms of minimization of prediction error because I've got some problems with but maybe maybe maybe maybe try to explain how I understand this if we take the few few level of description of David Maher I think that talking about the minimization of prediction error or minimization of uncertainty are related to the computational level and I think that talking about minimization of variational free energy or approximation by general inference are related to the algorithmic level and the idea which I try to develop is to using the connection between those levels to investigate what's happened on the mechanistic level the third level of Maher and I think that the notion of prediction error or surprise error when we talk about this we omit something important to what describe and explain Friston in his paper concerning the variational free energy yeah very interesting also there's different ways to view this but in the VFE centrism there's also the possibility of relationality as a core principle because if the mechanism surrounded suprisal of the system that would be closer to an account where there was like a surprise module or something that was about what the systems what representations were internal for the system but with VFE we have something that's explicitly implementational relative to a constructed model so then it it naturally connects to discussions about how modeling is done as a practice because it is a constructed variable whereas like temperature ostensibly it is important for a chemical reaction or temperatures of core screening I'm sure there's a lot of ways to say it more formally but the the kinetics do matter for the chemical reaction and then we have tools that are able to measure temperature in different settings cognitive modeling seems to have a relational step that's a lot more multifaceted because VFE is only calculated relative to the modelers construct so one person might make a model that shows the VFE doing this in a different model would show something different and that seems to be a bit different than like we have two thermometers but we both agree that temperature really does matter yes yes I agree I agree but I think the information that the dimension of information is most subtle than the dimension of temperature of mass and of course VFE is related to the modelers and model which they develop but I try to send something more of course VFE is related to the modelers but there is something in the target phenomena which motivate modelers to use the VFE and of course if I look on the piece of the paper the paper doesn't look like for example processes the information and it doesn't motivate me to to think about this paper as something what I can model and using the version of energy or version of principle but when I talk about on a living organism or the human brain on on the something like that there is something in the structure in the phenomena which motivate me to build a model then using this principle and of course the choice of the method of measurement the choice of the model method of model building is arbitrary in the sense that model the target phenomena doesn't directly give us an information which method is better but it's not arbitrary in the sense that I think that we can in the same fruitful way use the statistical mechanics for example to explain the brain and for example by Asian modeling to explain the brain it's not arbitrary in the sense that I can use the the piece of paper to to eat something as I can I can take a glass of the water if I want to drink but I cannot take a piece of paper if I want to if I wanted to drink but and this in in this sense when we distinct the former principle the aromatic principle we distinct the model which are built under normative principle and the target phenomena which we want to explain using this model based on the former principle then it means that that the relation between the target phenomena and the relation between the model is not arbitrary in the sense that we can use every kind of the model to every kind of phenomena and I think that when we try to look into our brains this kind of the structures in some way motivate us to build the computational models and those computational models I think is in some sense different than the computational model for example of behavior of no pigeons or moving of my hand and so in general I would say that choose of and coming back to your questions as I said the dimension of information that comes from international constraints are more more subtle than the constraint like mass or temperature because probably we don't have as same good method of measurement of the information like we have the good method of measurement of temperature of the mass and probably maybe the free energy principle or variational principles are not the best possible way of describing and explaining the brains but probably I think this is the most optimal for this moment when we take the whole knowledge our knowledge about the brains the most optimal models and methods to trying to explain this phenomena but but as I said using of the variational principles using of the Bayesian modeling doesn't doesn't don't lead us to the say something very specific ontologically independent from these models about the the the nature of the phenomena we describe thank you yeah that makes me think about a sufficient account a mechanistic account if you have a system that entirely is described by hot and cold flows through a pipe you may have a sufficient and even a total account with only thermodynamics now maybe the hot and the cold exactly signify like a signal so then there would need to be some kind of information processing model or layering to understand and and then we move into a space where um looking at just the information measures on the syntax like the amount of data transmitted in bytes or something over a digital connection you've already highlighted the relevance of the semantic information flows which is a topic that chris fields and others have started to look into and then there's a further uncoupling between the material basis of the information transfer and the semantic information flow i'll ask another question from the live chat so upcycle club wrote are there any limitations or challenges associated with incorporating variational free energy models into mechanisms yes good very very good questions um i think we can looking for the limitation related to the questions which scientist uh gives uh it means that we can looking for a limitation concerning the the the the methods and of course the more important limitation concerning the the border of the systems and it is hard to give a one good answer because in the field of them in the field of the new mechanical philosophy there is a big discussion about the border of the mechanisms and if we take the constraint-based mechanism approach the situation looks more complicated because system traditions of the composings mechanism will claim that the border of the mechanism are related to the components of the mechanism when i say that the components of the mechanism it means that something what we decompose is not enough because we have to also look into the elements which are not literally the part of the mechanism but which makes the mechanism active it means that the border of the mechanism is unclear and i think the main limitation concerning my idea is to is to say when we have to stop to explain for example when we take an example of the pipe and the water moving in the pipe this is the part of the mechanism but the the the the the material from which the pipe is built is the environmental constraints which is not the part of the mechanism but which is important which is relevant for the functioning of mechanism because the the plastic pipe if makes the better for example better working of the mechanism than the wooden pipe and i think the main limitation is related to the questions of which constraints for example when we talk about the constraints as environmental environmental factors which of the constraints are how many constraints we need to to say that okay this is enough we we understand how this mechanism is called active and i don't know for now where this border is because opening the mechanism on the constraints opening the mechanism on the information coming or processing it by the for example narrow mechanism make these questions about the border mechanics more specific or more more complicated but i think this is similar discussion or similar problem like we find in the field of the free nature principle when we talk about the border of the marco blankets or where we talk about the mark of the cognitive or mark of the living organism i think that if we integrate mechanically the free energy principle with the the for example mechanical predictive processing then then this discussion about the border of mechanism can be a part of the discussion of the border of marco blankets of the border informational systems like that thank you that's uh that's a huge point it made me think about having a physical object could be digital or analog that's going to implement some procedure program that is going to be bound apparently by the halting problem like the inability to know for some kinds of systems how the program is going to turn out by the incompleteness type patterns so there's certain informational constraints that kind of come out of nowhere seemingly to constrain the action or to to provide some some thing that limits the action or the know ability of the action so then you highlighted that it is actually this stopping time question or the halting question as a question of scientific strategy in building accounts knowing how much to include and and that's very related to like identifying or selecting a system of interest because you could say well the pipe comes from here and so we're going to go back up the supply chain and pretty soon if you include the flow of air or the movement of digital information or something like that in the system that that can unfold to include untenable models so then there's a core screening and already um people make informational core screenings like with causal influence diagrams like just the the the relevance of one variable on another and so it's just very interesting that by having the material thermo mechanism and a legitimated role or status for informational mechanisms which are the natural kind of explanation for perhaps algorithms or information processing that still there's the the boundedness of what is being accounted for but a clearer way to know how far out informationally just like you could go out physically in the thermal account with layers of rooms outside of the heat bath also you could go out informationally more into the past for example or you can just say we're just drawing a line around this so maybe yeah sorry sorry no I think that I think that the the the for maybe the best way is to um question about on the of the borders of the mechanism or borders of the system and relativize to the our scientific interesting because when we talk for example when we talk about the pipe we talk about the the digital system and there are some kind of artifacts and they are related to the architecture who made them when we talk about the brain it is hard to think about the brain in the same or analogous sense and we talk about the pipe or the glass or the paper because there are artifacts we can say about natural artifacts and I think that our thinking about the borders of the system we want to describe maybe primary should be related to the our our questions concerning the borders and probably maybe it the story was that that mechanists in the system tradition the traditional variable mechanism will say something like that we think about the the brains as mechanism like the other one we can we can develop the structure of the of the social web the structure of the of the brain the structure of the the the bones and so on and so on and this this way of thinking about the brain as the kind of the mechanics like the other one in some sense direct the investigation to brains and this is why for example decomposition as the whole strategy in your mechanical philosophy many authors think about this as as relevant and we don't need anything else but when Bechtel for example claims that okay but there is a mistake because our brains are based on the web of active mechanism and those active mechanism is most more specific because what makes them active is not only it's not it is not only what makes them and mechanism in traditional normal sense and when we talk when we look into the paper of Bechtel he doesn't develop the the idea of talking about the whole brain they developed the investigation into the small pieces of how some kinds of cells transform molecules and so on and so on but they they still claims that if we show the bound into too fast then then we don't explain correctly how the given mechanism the given cells do actively what does what actively what it does and I don't know if probably if I want to the defend the idea of moderate realism about Bayesian modeling I would like to say something like that that there are some kind of borders in the world and we tried looking for those borders and maybe this the systems and the difference between different systems motivate us to try and describe these borders but there is no possibility to say that the fact that here there is a borders but always when we talk about the borders this this talking about the borders is related to our our measurement tools and the the methods so I think that the small move in mechanical philosophy which I describe as the constraint based my hands approach and we can interpret it as as an example of the situation when the phenomena motivate us to change our scientific practice because back 20 or 30 years developed the new mechanical philosophy but in the last five or six years he claims okay but this old view was not perfect we have to change something and I think this is this is the argument for the moderate interpretation of the scientific practice phenomena in some sense because we develop the good method to investigate phenomena phenomena in some sense motivate us to change to develop the new methods and change our our scientific heuristic and so on so wow that's awesome one other thing it made me think about was the differences the the relevance or the difference between the thermodynamic and the informational accounts become very clear when studying decentralized systems that are spatially disaggregated because when there's something that's physically enclosed and and it has like a unitary physical component and a unitary informational component like a desktop computer or a single tissue then there is an undeniable overlap between those two systems which kind of host and relate to each other but then when there's a system that has other things amidst it then just thinking here about nest mates in the ant colony well then spatial enclosure and just inclusion of all the material components seems to clearly capture too much and so that seems to suggest a more informational role which takes on many rich areas from semiotics and all these other semantic questions not just again a reduction of the informational to the syntactic informational like how much energy or material was transferred because again that will just conflate that'll make it's a approach that won't lead to the best research agenda on mechanistic accounts. Interesting so I think that the question about semantic information is really important here but if we take a biogen ideas about the brain which inferred the information coming from the environment this is important this information is something more than the syntactic relation but I think what is important here is what Tristan's really important claims that I think that information changed everything in that sense that when you try to literally think about biogen approximation and so on then it means that everything we want to describe and explain as scientists is strictly related to the information we received you know of course this is very very simple but every every object we observe every object we want to describe is always related to the information we have about this object and it means that there's no something like we've got a naked fact naked state of affairs maybe ontologically there are but from the point of the view of the finite and light is like us every time when we receive when some information every time when we observe something and we try to develop a model of some kind of phenomena it's it's always related to the some kind of information what we have about this phenomena and I try to say something very metaphysical I think that the variational point of the view as we can say something like that at the bias point of the view show us that there is no distinct border between the the the object we try to map between the the language we try to use to map this object because I think there is something Maxwell Ramstad has a right when they claim that we still we try to describe the objects who behave like the that we try to describe the territory which behave like a map which describes this territory and of course this is not only the metaphor I think that everything what scientists try to do is to reconfigure is to identify the structures which are extender now to the structure of thinking structure of information in which we use I don't know I think that that that Bayesian view about human brains about the world lead us to the one very important questions about about the informational nature of the world I don't know because I speculate now okay but when we when we think when we take seriously this kind of thinking about the science about about the entities the most basic the most fundamental future of the world is not a mass it's not a temperature it's not an energy but an information and I think the information is something could change a difference and literally it means that if you want to develop a science we cannot ignore this very basic fundamental futures of the world why because the information of course I don't want to say it's something like from psychism on informational something like that but I will say that information is always for him for someone and I would like to say that from some some time when I was reading the papers of freestyle and so on I still think about why we should or why we shouldn't ignore the information in our investigation and I think that that Bayesian modeling give us a good tool maybe is a plea for the place of information in the in the science because but not only literally talking about information okay we know it's obvious that brains process information but information as an entity which making possible many many of the phenomena and I think that our brains in short we say our brains are the entities which are which are possible because there are information which makes this mechanism this kind of mechanism but not the other case sorry I think this is really speculative metaphysical but it's hard to catch this idea yeah there's there's a lot there certainly the unity of the scientific and the broader informational endeavor is very interesting and so thinking about that explicitly in terms of epistemic foraging seeking out which measurements or observations to accrue and that being kind of on a continuum of agency of experimental selection and there's other like a fixed camera does not have control over the sensory observations it's making it could have internal cognitive control but but having a unified way to talk about what it is that science or just different ways of knowing even outside of what we might call professional science today that those different manifestations are are part of a ecosystem of shared intelligence part of a shared fabric with other information processing as chemical elements are in circulation in a chemical ecosystem what is the status of the semantic informational connections in semantic ecosystems and people could have viewpoints like well there are no semantic ecosystems there's just syntactic ones and you know there's no syntactic ones there's just material ones I mean people can always go that reductionist path but to the point where any given person says okay I've kind of bottomed out on reductionism here or I'm looking for another perspective there is going to be the question of what the semantic account is and what its status is and does it articulate with a thermodynamic account do they have a zone of indistinguishability do they have a broadly overlapping region maybe it could be one way in one system in a different way in a different system but these will constitute like some of the fundamental philosophy questions that you identified in terms of real mechanistic accounts for intelligent systems and using active inference to develop any kind of account I think we cannot ignore the semantic and maybe the one of the most important things that Friston says is that semantic is important in that sense that it is not only the the medium which we use for example when we if I correctly understand the origins of the the Fristonian free energy principle it comes from the method of measurement of the brain activity and someone can say okay this is only the method and there is no any implication of course but I think that the Friston will say something more it says that there are phenomena in which semantics play and not only the of course there are phenomena in which semantics play an important role but we cannot ignore the semantic dimension in our practice of science building or model building it means that semantic is not something like the piece of the paper I can use the piece of the paper when I wrote a paper or I can use a word or a fix and something like that and this is not important what a material is used I think that many authors or maybe peoples think about the semantic as some kind of the medium like the kind of the paper I use I think that the free energy or active inference framework show us that the information or semantic is something what makes possible for example to do this to do a science or make or maybe the information semantic is something what makes possible for us a human human being and maybe like a thermodynamic free energy if we don't transform the the molecules from the environment we don't live but the question is why science talk many things about the thermodynamic processes and talk enough not enough about the informational processes I think that the Bayer-Germs modeling a freestone approach lead us to the conclusion that there are systems for which the transfer of informations is a same important like to transform process the thermodynamic energy in the argument which I try to develop this from the never computation show us that always there is a trade off there is a compromise there is a relation between information processing and thermodynamic work coast if there is some kind of trade off why we don't take any usage but probably it as I said information is something what is hard to be measured like temperature like mass and and so on and so on and I believe that personal principles give us a chance to try to develop this of course remembering about this that those principles are only there are those principles lead us to the tools which are only the more or better more or minor the approximation of the real real phenomena okay well wow do you have any closing thoughts or where is your research in this area going to go I think that I've got a grant proposal about this idea and a few days ago I've I received a reject in polish center and the one of them I think this the only one the reason why they reject my proposal is that I don't propose a using or apply this idea into the the real modeling framework so the next steps I and I think I agree because this is very speculative I propose some kind of heuristics and I believe that this heuristic can be applied in a real model building strategy and the next step but now I'm thinking is that to meet person or to make a competence concerning that and find the finding the way to how to taste this idea and in practice and and still there is two options and the first one is that okay this is the the speculative idea and we don't need to think about the mechanism in terms of information constraints and so on and so on and the second is that okay this is a correct idea but but maybe variational principle by Asian modeling is useless here because everything what you need is coming from the thermodynamic of formation or the classical theory from formation like and so on so I think there are all many open doors and I hope I will take away to find some of them thank you well it was a great presentation in discussion you're welcome back any time to share what that next step looks like I really appreciate this I think it certainly is a great benefit to the literature to connect the kinds of things that people want to say and do to now how we are saying and doing it thank you thank you Daniel thank you very much thanks for questions that was a very nice experience thank you excellent all right till next time good night bye