 Hello everyone and welcome to the Active Inference Lab. Today it is August 9th, 2021 and we're here in guest stream number 8.1 with Mark Miller and guests. So this is going to be an awesome presentation followed by a discussion. So if you have any questions during the presentation please put them in the live chat and we'll make sure to ask the panel at the end of the presentation. So thanks again to all of our great guests for coming on and please take it away. Thank you. Excellent. Thanks for that, Daniel. Okay, I'm just going to share my screen. Oh, can you turn on screen sharing for me on your side? It's under the advanced so at the bottom where it says share screen there should be a little arrow and click that. There should be an advanced feature and then it should say allow. Got it. Okay, go for it. Yeah, that's it. Can you just see my slide? Does that look good? Excellent. Great. Okay. Well, Daniel, thanks. Thanks so much, man, for the invitation to come and give a talk. I mean, we've been trying to get, we've been trying to find an opportunity to speak together for, you know, a couple of months now. Well, since I moved to Japan so I'm excited to be here and I'm excited to be talking about this topic which we're really excited about. It's a new paper that actually Julie and Keverstein who's here right now and myself and Eric have a preprint at right now in the paper title, title of Predictance of Happiness and Well-Being. We're also anticipating that Inez Hippolito and Lars and Ben Smith will be here this evening. Lars has had a slight emergency so he's on the road but we're hoping he'll be able to call in because actually the work that we feature in this paper, it builds on or it's really closely related with the research that Inez and Lars are doing so the idea was that we're going to be able to get, we're going to be able to really talk to them about their individual portion so hopefully they're on the stream by the time we hit that place and if it turns out that they're not, that's okay. We can just go through and then we can chat at the end. If you don't know me my name is Mark Miller. I'm an assistant professor at the University of Hokkaido in their Center for Human Nature Artificial Intelligence and Neuroscience and I am an active inference researcher. Right, so let's see. Right, so I guess probably lots of people who are listening to the stream tonight and definitely the guests who are here right now Hey Inez, great you made it right on. Know that recently there's been a huge amount of research coming out that's linking using the active inference framework in order to investigate all sorts of psychopathologies mostly finding a home in computational psychiatry circles but really we're seeing active inference accounts come up relating to depression and addiction and OCD and schizophrenia and dissociative disorders and all sorts and while there is this sort of active inference boom right now in relationship to computational psychiatry and in particular psychopathology we're not seeing any work really talking about how we might apply active inference frameworks to the theoretical and formal approach of discussing things like well-being so you know all those topics that would naturally fit in something like a positive psychology are still primarily being overlooked although you know in all fairness we are starting to see it come up at the edges because we're starting to see people talk about the benefits of hallucinogens and also the benefits of meditation so that was part of our motivation in writing this paper was to see if there was anything interesting to say in a positive account yeah not just the negative account but the positive account the other reason why we decided to write this paper was because actually Julian and Eric and I have written a series of papers now on psychopathologies in particular using the active inference framework so we have papers out now on depression and mood disorders and addiction and depersonalization and OCD and we started to see that there were some shared characteristics between these various psychopathologies if you look at them through the computational lens of active inference and seeing those shared characteristics sort of started providing we think clues about what might be going right when a human is happy, healthy and well and that's what our paper is about and that's what this guest stream will be about as well looking a little bit more at what some of the predictive dynamics might be that are underlying subjective happiness and overall well-being and of course it's a big topic so we're not trying to exhaust the whole of what well-being would be but rather just taking the sort of first step to see if we can work out some of the points that might be interesting so since I have these lovely people here tonight with me I've decided to make my presentation relatively short and I hope it won't be any more than sort of 30 minutes all said and I'm going to pitch it at a relatively high level of abstraction because I think there's like a number of nice moves I'd like to give to the audience here tonight to sort of provoke research and provoke some ideas and if you're interested more in the nitty-gritty we have the paper out in pre-print so you're more than welcome to go and actually check out the paper and of course if any questions come up we can talk about that as well great okay so I think a sort of natural place to start thinking about the relationship between active inference and well-being is to think about optimality in cognitive neuroscience and psychology today well-being is often spoke of as a form of optimal psychological functioning but what exactly optimal means is not very well explained usually or at least it's underdeveloped and of course as a free energy principle and active inference framework researcher when you look at that you the first thing I thought was well there's a natural way to sort of be thinking about optimality in terms of the active inference framework free energy minimization over the long run approximates optimal Bayesian inference is a form of optimal belief updating so that fact in some way already gives us you know it gives the free energy principle in the active inference framework you might think a sort of normativity already you know predictive systems tend to create or strive towards an optimal fit between their phenotypical self-organization and the sort of statistical regularities that they encounter in their environments so we think that an important part of well-being from this perspective is going to have something to do with being a good predictor of the sort of hidden causes so that's both in the environment and inside the agent of the sort of sensory states that we tend to encounter and of course that crucially relies on our being able to optimally update our generative model in ways that are life sustaining and flourishing inducing that that's an incredible stretch just as an opening gamut you know we know that well-being is closely related with health in various layers of health you know mental health and emotional health and physical health and good predictors they tend to self-organize in ways that allow them to maintain homeostasis across various contexts and over the long term so and you know and sort of first of all if we think along these lines which I think could be fruitful lines to think along then we have at our disposal as lots of the listeners today already know this sort of ever-evolving and ever-growing mathematical suite to help us describe what is a good model what's a good model all about so that's our starting point. Our starting point was to think well maybe well-being could be discussed broadly as a kind of optimal updating the goodness of the generative model that the agent uses to form beliefs and control their actions in life-sustaining ways over the long term. Okay so that's a good start and I think it fits very nicely with what we already know about lots of psychopathologies in particular that psychopathologies are characterized by sub-optimal beliefs and the sorts of behaviors that come from having sub-optimal beliefs. This is a good example it's just one out of the hat though really I think if you look at any psychopathology you're going to find some sort of bad belief is at play at some level of the hierarchy. I think depression is a nice example you know recent accounts of depression and not just from the active inference camp are highlighting the role that rigid beliefs about one's inability to manage the complexities of their life including especially the social complexities of their life is playing a major role in depressive symptoms and we've written a paper on this pretty recently and in that paper we using the active inference framework we argued that when an agent is met with enough persistent volatility then they can eventually develop a high level prior about their own high level prior belief about their own inability to manage volatility and when you get that high level prior installed pretty nasty can happen right the system a predictive system then starts selectively sampling the world to try to confirm that hypothesis and in fact it down regulates any sort of counter evidence that goes against that deep prior that deep pathological prior so that even when the situation changes you know if the agent finds themselves in more positively predictable situations they tend to hold on to that belief in volatility and that's really the pathological moment so another way of saying that is an agent at least a depressed agent yeah has become overly confident that it is uncertain it's overly confident about being uncertain we're not the only people who've said that clerking colleagues have written a fantastic paper definitely check it out if you don't know it and in that they come to the same conclusions they write in this sense we might conjecture that major depression occurs when the brain is certain that it will encounter an uncertain environment okay and as we all know you know due to the central role that prediction plays in sculpting our experience and directing our behaviors of course as soon as we have a high level bias like this installed then it has tremendous power to produce exactly the outcomes that would confirm those biases it's one of the problems of being a predictive agent you know the predictive system is really good for lots of things but one of the problems is there's this sort of deep self-fulfilling prophecy that can occur okay the rigidity of these bad beliefs is especially important here and it comes up again when we start thinking about what well-being might be you know there's mounting evidence that cognitive rigidity you know the sort of stickiness of bad beliefs that they're a basic feature of psychopathologies and in fact you know some people are even saying that it might be the root cause of all psychopathologies and I don't think we disagree there's a nice a on article about this if you if you're interested in that point and I think I might just take a take a little break there and hand it over to Anais, Anais Hippolito has a really just a fantastic forthcoming paper coming out that everyone should be excited to be reading which offers a really beautiful computationally rich account of how insulated internal states yeah when these bad beliefs sort of get stuck and then they resist updating how that creates the sorts of pathological belief systems the pathological behaviors that we see in psychopathology so if it's okay Anais I'm just going to stop share and let you maybe talk a little bit about insulated bad beliefs and then I'll pick it up from here and keep going on to well-being yeah is that okay yes for sure hi everyone can you hear me and yeah okay great so first of all thanks so much for a mark for inviting me to be here and it's always such a pleasure to be here with Daniel for these wonderful sessions and hi Julian very nice to see you as well so I'll just now share my screen if I think that's possible yeah let me see okay so for some reason I'm not able to share my screen I think it should maybe try again no it doesn't go through for some reason strange it works for me fine yeah I have no idea we could go for an audio thought or you can try to share it later let me just leave and come back if that's okay maybe that's good you know the trick this is the new world of technology isn't it you have to do everything on the fly you have to do everything on the fly and be willing to adapt I have one interim question yeah please yeah hit it were you working on active inference or interested in those approaches and then came to be studying psychopathology or just well-being more broadly or was it the other direction like which way did you know at this intersection from yeah you're right and it started the other way around so in my masters in which Julian was my advisor at the university of Edinburgh in 2010 the last chapter of my senior thesis was a proposal so I was working on empathy looking at you know philosophy and cognitive science of empathy looking at dynamic systems theories of emotion the last chapter was on predictive processing and active inference so I did my PhD with Andy Clark specifically on active inference and then it grew out of doing that into psychopathology yeah we got it very good we got it some technical difficulties but we are now good okay all right you can see it fine yeah yep nice work thanks and let's continue perfect okay so yeah so this work is on what we call psychotic mark of blankets and this is something that I've been working on for the last few months with two different groups applying this to different to different things so one specifically schizophrenia specifically and the other one is disembodiment so what I thought I would do here today be to do like sort of like combination between the two so that we get like a bigger picture of what we have in mind with this particular framework so yeah so let me just give some like framing about where this comes from and why and to motivate it a little bit so the part from this concept that Mark also just mentioned the self-organization which is super relevant to these things that we are talking about when we refer to well-being and mental health and all of that and as a few words about that an emergence of order from disorder and specifically without an intervention from the external intervention that is external to the system and then what is important here for us is to think of self-organization as a pattern or an arrangement that gives rise to a specific way in which the system is self-producing and self-distinguishing and here and not to the actual so but the idea here that the fundamental concept is the self-organizing pattern that comes up this is why why are we talking about that because we want to look at how the systems or the organisms as a whole are interacting with the environment in a way that has been called in modern literature as a tune with the environment right so the idea here is that these organisms or the systems are tuned with the environment and in a way that they will show a certain specific pattern or pattern arrangement that will be distinct from for example systems that are not tuned to the environment and this is very well known in the whole bulk of literature in disembodiment and phenomenology so the idea we want to pursue here is actually relating to a little bit to a complex systems theory in which we want to say that this differentiation between embodiment and disembodiment are different patterns it doesn't mean that the system that is not adjusted to the environment should be considered as disordered but instead that it has a different kind of order right so here we would be sort of like projecting a little bit the computationalist accounts of psychopathology as simply a disorder or an illness and opening up a bit of more like broad scope or spectrum to talk about mental health so then if we set up things like that then what needs to be explained is how the self-organizing behavior of situated organisms come to be ordered in a certain way rather than another and then we'll be now talking about mental health so then we think that active inference is a very interesting and useful tool precisely to think about this self-organization and the patterns that come from self-organization we think about active inference as a model of living things adjusted in the environment and we all know what active inference is so I'm just going to highlight what is relevant here for us in terms of pattern dynamics we think that it is a nice and cool model that when associated with tools such as marker blankets or renormalization group it can be applied to many different scales of self-organizing activity and that's where we find all of these patterns of our interest of our scientific interest so it is useful to explain these gradients spectrums of kinds of patterns of order in self-organizing systems so just highlighting these ideas so we take it as a model in this case so it means that we are not committing to the realist view of that phenomenon studied under active inference ontological process or leverages the conceptual tool but that the tool is really important and useful for us to understand certain phenomena in this case particularly what you want to understand is disembodiment right so we all know active inference but again let me just highlight just some important points to understand then the psychotic marker blankets later on so here the idea is that self-organizing self-producing self-distinguishing system will correspond to the internal states which affect and which affects and are affected by the environment and external states via a set of other states such as the active and the sensory states active and sensory states are then constructed as influencing one another giving rise to a certain part of self-organization and this is what's going to be relevant for later on so in a way what we want to say is that well an adjusted system that self-organizes is a system that minimizes the free energy which we all know can also be understood as uncertainty or entropy and so on right but an adjusted system would be something like that a system that is just well adjusted to its environment and for technicalities obviously there are these references on the bottom now moving on to disembodiment and then later on applying we want to apply this active inference strategy to further highlight or understand what is going on with disembodiment just a few words about the disembodiment literature so the idea is that comes from phenomenology and a little bit from existentialist philosophy and the idea is that the primary experience of the body is the body experienced as a subject and this is really crucial the thing is that this experience of the body as a subject can be lost in which case the body is inspected as an object more like a side of scientific interest and here is what the literature says that that's when disembodiment comes up is when the body all of a sudden seizes to be experienced as a subject and is now inspected as an object so what happens is in exploring and interacting with the local environment we usually typically we do not attend to our body we just like take it for granted we've got our body that allows us to access, move around, interact explore the world through our body and it is not until we cannot do these things or something is wrong that we will direct our awareness towards the body and then the body becomes salient as an object of inquiry and it's right here at this point where we find disembodiment it's when we cannot explore or interact with the world as one would like. So in a way we can distinguish embodiment and disembodiment by saying for example or just highlighting these features of course the literature is vast but I think that these are the ones that work for us here embodiment is an attunement with the environment that is allowed or enabled by having an experience of the body as a subject and that gets us attuned to the environment through a bodily self-awareness or something that there is a bodily self-awareness that is felt. Then disembodiment on the other hand is a disattunement with the environment because there is a loss of the body as a subject body becomes an object, there's a disattunement and a diminished body awareness. So this comes from the literature and it is interesting for us here why because then there's been these links that have been made between psychopathology and psychedelics and disembodiment as in disembodiment being something that is reported in different psychopathologies such as schizophrenia, personalization, even major depression or other psychopathologies that involve rumination. And another very important thing is that disembodiment also occurs or is reported in the in psychedelics intake and this has been even more relevant to studying disembodiment precisely in the psychedelic renaissance that we are witnessing today. And also another point that I want to make just here sort of like in brackets is just that disembodiment does not necessarily need to be a bad thing. We usually put things in good or bad right or wrong but it doesn't need to be a bad thing because you get some therapeutic effects, positive effects from psychedelics just precisely by having this disembodiment experience. So it's just a different way of interacting with the world and that's why we want to pursue less of a disordered way and much more of a there's a spectrum of different patterns of interaction with the world. Okay so just that not then I want to look at disembodiment and how we can use active inference to understand disembodiment in two ways. One is the psychology of disembodiment so how that disembodiment is felt psychologically and the other one is the neurobiology of disembodiment. So as we have said disembodiment is the loss of the body as a subject which is replaced by some form of intellectual effort. So imagine that you now don't have these broadly access to the world that is not possible because that's exactly what has been lost by psychology or psychedelics intake and many other things right. So then try to cope and compensate by literally engaging in making inferences about the state of affairs. So here we could use a little bit of the literature and sophisticated active inference which is this conscious reasoning and planning and policy selection that you engage now with having lost these access to the world through the body. Now our hypothesis is that the issue is that in this disembodiment because we all engage in sophisticated active inference that's all fine. The difficulty is that in the case of disembodiment experience this inference is isolated, is isolated because what has been lost is the broadly access to the world. So then what happens is so we would have something that looks like this. So we all know this scheme here right and what we want to say is that what is lost now is the influences between sensory and active states and let me guide you through it. So let's just focus on the figure for now. So we have internal states as being disembodied reasoning so a subject that has lost this access or broadly access to the world because the body is not felt as a subject. So we've got this disembodied reasoning as internal states and then we have external states as the environment where that the subject wants to explore and interact with. And then we have a sensory states, the environmental co-constraints or these aspects that should influence the subject. And then as active states we have the affectivity as the ways in which the subject can affect and be affected by the environment. But what has been lost here is that there is an imbalance between sensory and active states. So for one, active states are not sensitive to sensory states. So it is this affectivity or this possibility of being affected by the environment that is not possible. Why? Because the access through the body as a subject has been lost in this psychopatology or psychedelics as we label as disembodiment experience. So as a coping mechanism what is going to happen is there will be an increase of activity of the activity of active states by increasing the issuing of inferences about the world. So I do not understand what is going on, but I'm going to continue to think about it and make and develop theories about what is going on. The problem is that we all engage in these developing theories and inferences about what we think is the likelihood of the state affairs. The situation here in disembodiment is that these inferences are disembodied. So they tend to be false beliefs and that's because of the insulation of internal states. What does this mean? This means that there is not to the access that we had with the world through our lived body is what is lost. So then the subject in a situation where the only hope is to increase active states to make more inferences about the world to cope with that loss and then tendency is to have false beliefs. That's when we start entering the realm of psychopatology and psychosis. So in these disembodied cases this is actually a very important point because one could say, okay, so then the problem is on the level of reasoning. So on the level of for example the internal states as we have set up here. No, the problem is not on the level of the internal states. Reasoning is fine. The inference mechanisms are working perfectly fine. The subject can engage in reasoning. It can make inferences. The problem is inside the Markov blanket. The problem is in the unbalance that there is between active and sensory states because those are the ones that are supposed to directly influence one other such that indirectly there is a connection between internal and external states. But it is this direct influence between sensory and active states that is problematic. They are not communicating well. They are not influencing themselves or each other in a balanced way. And in a sense, in balancing the sense that active states become super active and sensory states cannot influence active states. So then what we get to this idea of a psychotic Markov linkage which is if we have this sort of arrangement where there is a problem between the balance of the strength of influences between sensory and active states, then we have a problem of the generation of false beliefs. The problem being that these beliefs that are issued by active states they cannot be updated by the world because it is precisely this access to the world that has been lost, lost by virtue of having lost the body as a subject. So then because of this attunement that we can actually explain as what is occurring inside the Markov blanket, the reasoning subject makes inferences about the world, but this reasoning occurs from within the lack of the experience of the body as a subject in an insulated manner. And this is consistent actually with a long-standing literature and phenomenology of psychosis where the subject, there is a bunch of literature on psychosis linked to solipsistic behavior, solipsistic drawing and painting, and also the philosophical idea, which is the philosophical idea that only one's mind is sure to exist. And then in the reports that I find quite interesting, some patients would report, I was no longer sure that I was still the same person. I became one with other creatures or objects. I lost a sense of my own boundaries. When I had an experience, I often did not know it was mine or the experience of someone else. So it is quite interesting what can be lost when we lose this body as a subject as it is developed in phenomenology. This also happens a little bit in the personalization disorder, for example, showing this disembodiment to the tournament when making people feel estranged and detached from their self-body and the world. Okay, so this is all I had on the psychology of disembodiment and now I would like to turn to the neurobiology and also in the neurobiology and apply the same framework now to the neurobiology of the underlying neurobiology of disembodiment. So we think that applying the same reasoning is consistent. So this disattunement that we're talking about felt at an agent level should also be expected at a neurological state. So that's our hypothesis. So now we think that there are two good candidates, would be good candidates, so proprioception as a system underlying movement and location of body in the environment as well as interception as the system responsible for homeostasis. These are good candidates we think for the neurobiology of this general disattunement that is felt in disembodiment. So proprioceptive and interceptive systems just like any situated organism can be conceived as self-organizing systems so we apply exactly the same principle as we would apply to the agent as a self-organizing system. So the question then that we had is what is the factor that is at the basis of this neurobiological pattern underlying proprioception and interception that will show this maladjustment to the local environment. So if we expect the same sort of mechanism or organization or dynamics as to the whole agent that we were just talking about then how are we going to find this out as well in the neurobiology So we think that just like in psychological experience of disembodiment where the agent becomes insulated from the world and in this case we say by virtual losing the body as a subject the answer will reside in biological installation. Now I'm going to focus now on interception because I don't have time to also cover proprioception and I think that this does give us here some very very cool insights. So I'm going to focus now on the work brought by Barrett and Siemens on interceptive predictions in the brain and they have a very cool model which they call embodied predictive interception coding model and what they goal there is to integrate an anatomical model or cortical connections with Bayesian active inference principles and they propose a very cool thing which is that a granular visceral motor cortices contribute to interception by issuing interception predictions So more specifically what they want to say is that the granular limbic regions that regulate visceral motor of the body's internal milieu that is visceral motor cortices so this is their focus and where they think they will find the interceptive system and where they applied the model. Now so you can see the agranular cortex here on the left and one very interesting thing that I'm just going to draw your attention to and then if you feel that you'd like to read this super interesting paper then please go for it but I just want to draw your attention to the fact that on the left side you have the agranular cortex and there's one thing that is missing from the agranular cortex as opposed to the dis granular and the granular cortices so what is missing is the layer 4 and what happens with that is that there won't be you cannot see there any prediction error neurons so that's pretty cool because that will give us an idea that there is no prediction error neurons in the agranular cortex so just I'm just going to leave that out there and just moving forward to just like an understanding like the very gist of the embodied predictive interceptive coding model they claim that there is an interceptive system in the brain in which agranular cortices send visceral motor predictions to the body and transmit interceptive predictions about the visceral sensory consequences of those predictions so they also tell us that agranular visceral motor cortices estimate the balance between the autonomic, metabolic and immunological resources that are available to the body and the predicted requirements of the body based on past experiments right so this is going to be very important for us because a chronic imbalance of those things that I was just mentioning can be which can be caused by constantly predicting the need for more metabolic energy to demands for stressors can produce the well-known depression-related disruption and we can also eventually down-regulate, provoke a down-regulation of hypothalamus or HPA axis negative feedback loops resulting in chronic hypercortisolemia and in turn can also promote pro-informatory state so this is like some of the problems of having an unbalanced on the level of the autonomic, metabolic and immunological resources which is what the interceptive system should be able to do is to give us that nice balance right so then applying exactly the same framework, the same model that we applied to the whole body, disembodiment of the generalized feel of disembodiment that is vastly described in the phenomenology literature, applying this to neurobiology of the interceptive system then what we get is something that can be seemed as like quite similar an unbalance between sensory and active states so let me just guide through the figure so here in this figure internal states are or correspond to the interceptive system and the external states, the environment whichever the multi-scaled system the interceptive system is part of and then we have active states as the agranular cortices that send visceral motor predictions and then we have sensory states, the requirements of the body given the history that the body knows to be required for human states so then in this sense what we would have is that active states are not sensitive to sensory states so even though active states can become as a coping mechanism much more active and send much more predictions about what is going on, what's the state of affairs because I'm not receiving any information from the requirements of the body so the sensory states so what will happen then because of this insulation is that a false belief is going to be generated so this will correspond to a constantly prediction of the need for more metabolic energy to meet the demands of the stressors so in this sense we have here another case of what we call a psychotic Markov blanket, why is it psychotic? because it delivers if there is this lack or disembalance between sensory states or the influences between sensory states and active states then what we should expect is a degeneration of a false belief and we can also expect that at a neurobiological level which is also supported by empirical evidence so yeah this is what I wanted to say just thinking these collaborators here that have been thinking about and guiding me through as well these ideas. Thank you Ness, really interesting stuff so I guess we'll return to Mark and anybody is, yep anyone's free to ask a question in the live chat and I'm writing things down and then after this second piece of the presentation we'll have a discussion so thanks again go for it Mark. Yeah that was great thank you so much wow it's so interesting and I haven't heard some of your stuff about disembodiment yet and you know also this is a research program that I'm extremely interested in so I'll be picking your brain at the end of the Q&A and also after what might, what wasn't made obvious in Ness's talk but I think is implied is that the stickiness of these belief systems and what's helping them become insulated is a story about how precision is being estimated and how precision is being deployed that's surely one of the mechanisms that is creating an imbalance between active and sensory states so you know how precision is estimated is really the ability to assess the uncertainty of its own priors and when that's working right then those estimation should result in appropriate model updating you know you shouldn't get stuck in an insulated situation but rather you should know wow that's not working out the right way and so appropriate updating should occur and we know that like maladaptive precision estimation is disastrous right and we're seeing it everywhere in the psychopathology literature today you know schizophrenic delusion is being characterized as too little precision on sensory data which is putting this heavy emphasis on just top-down predictions so that's why you're seeing the world in a way that isn't being anchored by what's really happening and of course ASD is a little bit the flip right too much precision on prediction errors creates a lot of volatility and a little bit of difficulty establishing good high level predictions and as a quick side note you know the kind of insulated outcomes that create these maladaptive self-organizing styles they can occur in lots of ways right so for instance yeah so addictive substances is one good way right addictive substances impact directly precision estimation machinery right in the way that the addictive substance engages the dopaminergic system so addictive substances are setting precision outright in a way so no wonder they can create these sort of sub-optimal styles but societal expectations do the same sort of thing you know sub-optimally pinning precision on expectations about success for example you know so you can have cultural pinning of precision that you should achieve a certain level of success in your life and you know we have deep running relationships between perfectionism and depression for example and again for exactly the same reason that and as said overly high precision on some sort of expectation would lead to a constant pervasive background error as you're not hitting that high level of expectation and then just like Feldman Barrett suggests in Simmons you have an HPA flip because the system just doesn't have unlimited resources and then the one that I'm most interested in today actually is social media which is turning out to be a like fantastic way of bending our generative model into all sorts of shapes that cause sub-optimal outcomes right and that makes sense it's just giving us wave after wave of you know misinformation through filtered images and staged images and of course social media platforms some social media platforms function just like addictive substances or like casinos in a way so again they're hijacking the precision system in just the sorts of ways that cause these problems I just want to quickly summarize before I go on to why what I think the natural mechanisms that keep us out of these problems are so well being as optimal model generation importantly depends on correct and flexible yeah not these stuck insulated outcomes but correct and flexible self-estimations of uncertainty so that's precision assignments so misassigning precision to prediction errors which can happen in lots of ways especially in our modern world will inevitably lead to sub-optimal styles of self-organization sub-optimal generative models and with that the production of pathological beliefs and pathological behaviors and so that's how it can go wrong or at least this is the starting point it's one of the ways that it can go wrong so that's how it goes wrong then what is it that ordinarily keeps us out of these sorts of bad bootstraps these sorts of insulated disharmonies between the system and the environment and one answer that I like that we like is a team it was given by Alex Chance and colleagues where they said that optimal free energy minimizing agents will benefit from striking a good balance between pragmatic and epistemic engagements okay so that is they don't only act in a pragmatic way that is in ways that they're highly confident will reduce free energy over time because they can be wrong that's part of the take-home message from some of the pathological self-organization styles right they say that you should also a good free energy minimizing system is going to balance that against being driven to epistemically explore and forage for more information about how the world works but also about how their own model works and implied there is that when they find important counter-evidence while they're out exploring that they're willing to update their model relative to that counter-evidence so then the question then if we buy that and I think I do buy that it's one of the ways that we stay out of these bad bootstraps sub-optimal styles of interacting well how do we find that right balance I think that's a good question if we're going to ask well how do we live well what is well being all about how do we find that good balance and the central way that we've been arguing over a number of papers is that we literally feel our way to that balance and that part of course is predicated on everything working right because if you have an affective disorder then it's going to be exactly that mechanism that's going to be mistuned and of course we get all the problems that Anais just pointed out. So in a series of recent papers we've been exploring the suggestion that good precision assignment is going to rely in part on the rate at which error is being reduced so that's commonly referred to as aerodynamics and we're increasingly not alone in this conjecture we have people like Geoffrey and Courcelli who gave the first paper on this and of course Vandekruz and then it was beautifully formalized by Casper Hesse and colleagues. So if you want to look really at the nitty-gritty of this that's definitely the paper to check out and if you want to see a great paper on the computational bits of this you should see Casper's talk it's there on the slide it's fantastic. So the idea the idea has developed that these aerodynamics the changes in the rate at which error is reduced are registered by the organism as affective states and if we think of an agent's performance in reducing error in terms of a slope that plots the various speeds that prediction error is being accommodated relative to our expectations then we can think that positively and negatively balanced affective states are a reflection of better than or worse than slopes of error reduction. So that is just to say we feel good when we're doing better than expected at reducing error relative to expectations, better than expected we feel bad when we do worse than expected at reducing prediction error and these embodied aerodynamics are important drivers of precision estimation it turns out and I think that makes sense of course right because unexpected increases or decreases in volatility relative to your model is important information for how you should set precision on your beliefs about action policies right so unexpected decreases in the rate at which error is being reduced tells you that your action policy isn't working the way that you thought it was going to work which means you should obviously reduce your confidence in that action policy and the other way as well right if you're unexpectedly doing better at reducing prediction error relative to some policy then you should up your confidence in that policy so the point being here is that precision then isn't adjusted only based on the amount of error reduction but precision is also being tuned relative to the rates at which error is being minimized over time and that feature of the active inference framework gives us a rather sexy account of momentary subjective happiness which I think is so interesting right momentary subjective happiness is the result of unexpectedly reducing prediction error and that feels good because we are suddenly doing better than we thought we would at getting a good predictive grip on our environment which under normal circumstances would usually mean that we are staying healthy and we're staying well and I'm Jeff Lee in coracelli's graph makes that clear you know as free energy is going up you feel fear and then you feel unhappy when it's super high and you feel hope when it starts getting better and you feel happy when error is being managed in an extremely good way but certainly showing up an active inference these ideas are being presented in other camps as well Rob Rutledge is of course the paradigmatic case he and his team have done a number of brain imaging studies that have already started to show the strong relationship between subjective feelings of happiness and doing better than expected in some particular niche you can see on the board behind Rob there that's the calculation for happiness which I think is a bit of a cheeky way to go but I think it's cool because if you look at the math then of course you're going to know what's going on there too okay now we get to the fun stuff this is the last thing this is sort of we're coming into the last stretch here's the fun stuff so we think that the same machinery that underlies momentary subjective happiness is also going to play a leading role in helping us tend towards optimality and to avoid these sort of suboptimal styles of self-organization so we think it's going to play a special role in an active to see why aerodynamics play that role we just have to ask ourselves well where are the best slopes to be found and the answer is the best slopes are found at the edge of our own predictive capabilities right between what's known and highly reliable and what's unknown and potentially more optimal and when it's all working right aerodynamics naturally move us out to that edge right where errors are neither so complex that we can't learn anything from them nor too easily predictable that there's no more epistemic value and really the best slope is going to be right there at the edge of your own capabilities and you know that's not a provocative idea we see that everywhere in the curiosity literature we know that babies are attracted to medium amounts of complexity not too well known not too volatile we see it in mice put them in a maze with varying rooms of grades of complexity and the mouse will tend to hang out in the room that's just above its comfortability and you know we're already building adaptive and effective robots using these ideas so this isn't a far out idea so important for our current discussion is that in order to make the most of those good slopes that are right at the edge of our own capability predictive system must be sometimes willing to disrupt its own fixed point attractors yeah so it must be willing to go beyond its habitual policies to go beyond it's extremely well known ways of self organizing beyond its own homeostatic set points even right and that's one of the roles that aerodynamics should play when everything is working right when it's going well and you're on a good slope of error reduction then you should be motivated to continue along that path but when the niche thesis to give you productive prediction errors then the negative valence that comes up that signals to the system that it's time to destroy some of these fixed point attractors in favor of more wandering policies more explorative styles and I mean if you want to see that in technical terms just check out Friston, Brakespear and Deco that's the paper to go to so agents that use aerodynamics in this way to set precision on action policies they're going to avoid getting stuck in any one fixed attractor basin for too long right they'll instead tend to exist in more metastable states more often they're going to be poised they're going to be optimally poised in some way between exploiting already existing off action policies and performing action information seeking epistemic actions that reduce uncertainty and that metastable poise actually grants all sorts of fitness advantages over strictly ordered or more chaotic systems at least in our kind of environment and that's something that's very well known in the dynamic systems literature so again this isn't a very provocative claim here and it's precisely because those sorts of systems have an optimal balance between you know what you might think of as efficiency and degeneracy another way of saying that is they're set up to respond efficiently to particular contexts those contexts that are highly predictable but they're also remaining open to exploring and growing in a wide variety of other possibilities okay so notice then that for a free energy minimizing system in our kind of environment flourishing is not about avoiding errors I know that's starting to be more peculiar but I think it's a really good point to like hit right on the head it doesn't flourish by avoiding all errors it flourishes by selectively sampling the right errors I think that's so cool like if that wasn't immediately obvious if you're still not like you're still not really catching that point and I think it's a good point to catch soon what matters for flourishing for a free energy minimizing system in our kind of environment is that it gets the right kinds of errors into it and of course we know that you know like outside of the model it only sounds weird I think inside the model if you're in that camp outside of the model we already know that you know overly sterilized environments make children sick overly simplified cities make people depressed like we know too little volatility causes problems for complex self-organizing systems like us so again I don't think it's such a provocative claim although it can sound provocative from inside the active inference framework maybe so finding the right sorts of volatility and the sweet of cognitive emotional mechanisms that keep us in touch with that good volatility we think that that's going to play center stage in any account of well-being if we're going to look at it through the active inference lens and we've done a little bit of work on this already this is our aon article where we do this in a sort of popular way if you want to check it out the title of the article is the value of uncertainty good okay last little bit and then we're done so in and now we're going to come into the most speculative part of the idea and so then we'll stop and then we can talk about it and see where it goes so in the literature on well-being researchers tend to congregate in a couple of big camps and two of the most popular camps are thinking about well-being in terms of feeling good that's the hadonic camp and people who think that well-being is about being good and that's the eudaimoneic camp and actually those go all the way back to Aristotle featured here as the cool guy in the glasses so we speculate in our paper that actually we might be able to map those two subjective momentary subjective happiness and overall flourishing being a good sort of predictor on two forms of aerodynamics a local form of aerodynamics and a global form of aerodynamics and we think that there's a good reason to start thinking about how we might distinguish between local and global aerodynamics because for instance think about video game design okay so a well-crafted video game could be highly tailored in such a way that it provides you with just the right evolving changes to keep you continually making learning progress relative to the game so you're going to keep finding those good slopes and we all know good game design does do that it keeps creating these good slopes of air reduction so do good casinos right but just like with substance addiction right you can get so immersed in those local good slopes that you end up surfing them at the expense of the other things that matter in your life so you start neglecting your friendships and your school work and your overall fitness okay so while you're succeeding locally at the game you're not flourishing overall because in fact there's going to be rising volatility across all of these other domains that you're temporarily not paying attention to right so sensitivity to global aerodynamics would essentially amount to being driven in ways that reflect how well we're managing errors across the various domains of our life and as long as an agent was using those more global estimations of how well or poorly they're doing over their various concerns to adjust the precision then you're going to have an agent who tends to stay in touch and who tends to move towards optimality relative to those multiple cares and concerns so you're not only improving locally but we think that there's good reason to think that we're also sensitive to how we're improving overall and that's going to play a special role in well-being in particular and that makes if that's right if we're right here then that makes a really natural fit between this computational work that we've been interested in and more recent so-called network accounts of well-being where well-being is thought of as the result of tending to a complex ecosystem of positive states and positive achievements so then global aerodynamics would be crucial here we think to staying in touch with those various concerns in the kinds of way that would allow you as a whole to tend towards a good overall optimal shape now like I said this is the most speculative part of the talk nobody has talked yet about thinking about aerodynamics happening in a nested way but I just want to say that we do have good reason to be thinking along that line it's unfortunate that Lars isn't here with us tonight because it piggybacks on some of his deep parametric modeling so I was hoping to lay it off on him to really get into the computational nitty gritty of why we're not just pulling this out of a hat for you but we so we build up this idea that he's already been developing and the idea is basically he uses a deep parametric model and you can check out his paper on metacognition if there's a way that we can link this somewhere I can send it to you afterwards but he ends up showing that you know there's higher level policies that are governing the allocation of precision on lower level tasks and the kind of cool thing about this deep parametric model is that it becomes possible to appreciate precision over those higher level policies over precision allocation to particular tasks you get this sort of nested you get this nested precision estimation where at the lower level you have precision being adjusted relative to a particular task at a higher level you have precision adjustments that are also affective we suspect being set over the policies about how you deploy your precision over all the things you care about so that that better than or worse than the global better than or worse than is going to be all about are you doing better than or worse than at deploying precision across your various concerns not relative to some local concern so then we think that what breaks down an addiction in particular is you start paying attention to aerodynamics locally in the drug seeking and taking behaviors and you actually have a separation of context and we see that in the brain right you see the striatum pull apart from the dorsal lateral prefrontal cortex so the more contextualizing parts of the brain are no longer communicating in the same way as they were before because you have ultra high precision on these local aerodynamics so there again is a good example where we're going to get an insulated sort of psychotic Markov blanket emerge because of an overestimation of precision because the drugs of addiction are making that happen so Lars and myself are currently working on clarifying that aspect of the model so look forward to work soon on that and we're looking at how we might apply that model to thinking about meditative development thinking about attending to attention on precision modeling precision and thinking about how those sorts of mechanisms might help us have better access to and be better guided by these higher level precision tuning processes yeah so look forward to that and maybe that'll be another thing we can come on and talk about here on the lab and then hopefully Lars will be not an emergency situation that time and he can come on great so that's probably enough for me maybe just as a final note before we turn it over to Q&A I might just say the future directions for this are plenty if you buy some of the mechanisms that we've been thinking about right it would be cool to start now thinking about what sorts of endeavors encourage or disrupt that sort of metastable poise right meditation hallucinogens provocative art installations you know are they are they helping to understand during can we foster these metastable states you know and how so deep brain stimulation looks like it's doing something like this you're getting stuck in a tractor basin and OCD deep brain stimulation and Julian can talk about that if it's an interesting point reopens up the field of affordances by creating more dynamic neural ecosystems what does this mean for how we structure our environment if this is right the edge of our own capabilities and what does that mean for how we structure our environment how should we build our cities I mean Abby Taber is doing incredible work on this thinking about how we might structure our cities in ways that are good for us and actually they turn out to be sort of paradoxically not making the city easier for us but making it more challenging getting more diversity more art installations more gorilla gardening things because actually the more complexity we have actually the better we are as individuals and of course the biggest thing I'm working on right now here at Hokkaido University is what does that mean for how we design our technology and how we consume our technology how about you know what role is misinformation as a digital crowbar prying us apart from the statistical realities of our environment what role is that playing if these flexible fluid dynamics are part of what it is to be well that's all that's all future stuff and if you're listening to this tonight and you think it's a cool talk and there's something of value here then that's the way I suggest starting to think next and I'd love to hear anybody you know here on Q&A but also my door is always open if you sort of dig the talk and you dig the research then definitely reach out because we're always up for collaboration okay that's probably good for me you're welcome awesome so fun and so many cool ideas in there anyone in the chat is welcome to provide a question and first I'll just pass to Julian to maybe give any remarks or thoughts and then we'll see if there's any questions in the chat or some questions I've written down no I think Mark did a great job so let's just jump straight into the Q&A thanks Matt awesome okay so you mentioned it kind of tantalizingly at the end about how we would be using the technology and that just made me wonder like what would inactive inference social media look like or how would somebody of course not see any examples of it out in the wild how would we even go about imagining that kind of a system anybody else want to try to answer first because I've got something to say if anything comes up should I just go yeah I mean it's a good question something I'm really interested in maybe one thing to say is you might think that there's a big market for making tech that reduces our option set you know that oh well this tech is great for us because it simplifies our affordance landscape and that's what people want they want to sort of simplify it affordance landscape and the danger is that actually we're going to be attracted to that because it's going to temporarily reduce free energy faster than we thought it would right Google does a great job of quickly minimizing our uncertainty about various topics but over time you can have a sort of nasty outcome and I think we're starting to see that more and more where you collapse because it's so rewarding to come to one place like your smartphone ends up being one place to come again and again and again and over time your affordance landscape is collapsing it's increasingly collapsing to one and only one end it's a little bit like Walmart you want to buy there because you can get great deals but the problem is is that when the whole community buys there and all the mom and shop complexities go out of business then you're stuck just with Walmart and then they put up their prices it's a little bit you know it's sort of like that you think that the collapsing tech is going to be good for us but actually over time having less complexity is harmful for us and think about social media platforms as well like you sterilize social interaction and it's really rewarding at first because social interaction is hard it's a complex affair you know it's very uncertain how to manage it to be with another human so temporarily online interactions can be really really rewarding because they're so much simplified but if you hang out there for too long and you lose some of your ability to engage with complex human affairs then we're starting to see some of the ramifications of that sort of technology I mean we're literally dying of loneliness even though we're more superficially connected than we've ever been and I think that would be an interesting project is to start thinking about what is working when we have complex ecosystems of social interactions that we're losing by just using social media platforms for our social interaction yeah Julian yeah so I was thinking about intellectual complexity as this is something we've been talking about before and how important it is to be open to prediction errors if you're going to get the most out of social media so you can think of filter bubbles as a kind of psychotic Markov blanket in a way in that there is insulation from any evidence that doesn't match with your predictions or match with your expectations so in order to flourish to live well with social media perhaps what's necessary is this kind of openness to other contradictory points of view what the ancients would have thought of as humility and we need to foster that to be well with technologies so Ines what do you think about filter bubbles as an example of a psychotic Markov blanket do you think that's a stretch or does it fit with your the way that you describe this insulation from sensory evidence that happens that can induce a kind of solipsism right yeah absolutely I think it's exactly on the right track actually the way that algorithms are constructed to deliver precisely these bubbles in social media it's so interesting because it's precisely what's going to happen a psychotic Markov blanket because it's just going to be insulated from that amount of information making inferences about that precise information without having any other contact or exchange outside of the bubble so that's very interesting as well and there are many many other forms of psychotic Markov blankets that you can find in societies and communities and that kind of thing that's one of the exciting things about the concept that it's scale free so you see it at collective levels of self organization not just but also at the level of the cell you show so the NMDA receptor you can also think of in terms of a psychotic Markov blanket yes I didn't have time to present that here because yeah but that's also formulated under the disconnection hypothesis by Karl so it's also really cool because it's at a cellular level so it's quite interesting if anyone is interested so that would be also what we are applying or looking at in this paper as another example of a psychotic Markov blanket where there's an imbalance between again what I was saying the sensory and the active states formulated there as the receptor and specifically for schizophrenia but then you can scale all of these up and find all of these psychotic Markov blankets in many different levels and I think that in the case of social media the bubbles are an excellent example but I think that we also have that as Mark was mentioning and I always find it's very interesting when we discuss this topic about specifically the city living as an environmental daily living space influence mental health and specifically why and when some groups of individuals thrive in an urban setting so this is something that some people here in Amsterdam for example are working on and it's really interesting and I think that it becomes even more interesting if we come for example from studies in neurodiversity for example in autism and that kind of thing and if we think about the design of these spaces daily spaces daily living spaces that are not designed for example for diverse ways of interacting with the world such as on autistic interacting with the world so I think that it is very relevant to think about these mental problems and how we can inform interventions in the city planning and also smart environments that would allow a more diverse way of interacting with the environment as opposed to reducing the complexity of the environment which is exactly what Mark was talking about and I just wanted to add that to that. Yeah we're actually working with a pretty cool group through the nested minds network developing models of antifragility and what we're calling tropophilia which is the love of uncertainty so you're not just growing from volatility but you're actually seeking volatility in order to grow and we're thinking about it at the level of communities so we're still just getting started but it's a really fascinating field yeah. So one comment on that and then a question from the chat Julian you pointed out that active inference as a scale and sort of a base system agnostic framework really helps draw out these similarities across systems and then Mark in the presentation you talked about neural ecosystems, about social, social media ecosystems and then thinking about some of these pathologies or perceived pathologies in an ecosystem case I mean imagine if someone said I like this tree this should be the only tree or I'm seeing a little tension here with the predator and the prey we should separate those two species those would lead to a catastrophic collapse because it wouldn't be the fragile ecosystem that was the one that had made it all the way to the present moment and was existing far from equilibrium so it's sort of like when we can look across systems it gives us a new way to look at the systems that we can modify the niche of like our cities the internet and then starts hopefully shining some light on where we can look. Here's the question from the chat Julian wrote for Mark a simple provisioning rule when in doubt zoom in comma zoom out so that's Dean's providing that a sort of like a little simple motto as an attending strategy could this be a happy person setting up a local or global non isolating Markov blanket like how do we go from these mathematical framings and some of these aerodynamics into nice slogans like Dean does. No I love that I mean right yeah exactly yeah look close and look far I mean that's that sounds like that sounds like super good advice a little bit you know this a little bit follows on from the same point Julian was making earlier about virtue about some of like what would the virtues be relative to this kind of good updating system and I've been meaning to do I've been meaning to write something on this for a long time now and you know I'm getting more and more motivated as we go along but have you ever heard of super forecasters you heard of this it's a book I would call it super forecasters yeah so there's people around the world who go to competitions and they predict things from noisy data sets and you can train in it and you get in a team you can do individual events and people can get better so you can train to be a super forecaster so you take noisy data you get good at predicting and I like this idea because I think it's a nice metaphor you know I think we might be able to glean like if we're all predictors and we're trying to predict well then maybe there's something to learn from people who try to predict well as a hobby you know and try to see like what are they doing right and there's some really interesting things that they do that I think are useful given the kind of framework that we're proposing here a couple of them I think just everybody should know about one a super forecaster across the board the best the best super forecasters in the world they share a couple of common traits they are curious so they're always interested not just in their own field but they're interested in lots of field they tend to be polymaths so they're interested in lots of different sorts of things that's one thing the second thing and this is a little bit I feel like is related to this zoom in zoom out point and also exactly what Julian said is that they have they hold sort of this active inference language but they hold high level beliefs that all of their beliefs are only relatively right so they know that when they're looking at something not to get too enamored by it because it's up for updating at any time so they're always checking it against the other things they know they're the kind of person that when you meet them in public and you tell them something they don't know they don't lock down and say no no I can't be like that no no no you know they're the kind of person who says wow wow I didn't know that tell me more about that so I can update I can update my overall model relative to how the world is anyway I think I think these like this is zooming in and zooming out sounds just right and I think it's so cool to start thinking about what some what are some of the practical virtues that would help us not get stuck and I think I think what you just described to the super book houses is something that goes back to the ancients again the Peronian skeptics you put forward the idea that we shouldn't endorse a proposition but not should we withhold our endorsement we should rather try to find this balance and what you're describing with the super forecasters with that kind of hype I think is exactly that same kind of balance well that's the paper Peronian skepticism through the lens of active influence yeah in the ancient realm it made me think about skepticism and like almost like a Bayesian stoicism like the info has come in through the channel it came in from I would like to learn more I will marginally update like that sort of approach and then also another dialectic that you brought up that it'd be interesting to hear a little more about is that hedonic short term happiness versus that the eudaimonia the longer term happiness just like that's super fascinating to see how you mapped like two pillars of philosophy really going so far back and integrate them in a new way like where do we go from there how do we think about these kinds of cool synthesis between philosophy and computational dynamics I love that Julie do you want to say anything on that if not I could maybe start and you could follow on what do you think yeah so the way we were thinking about it was to take I don't know if Mark did this example in the talk to pop out but the computer game examples are you going through that mark yes you might you might find a computer game that gives you just the right amount of challenges to keep you constantly making progress in playing the game but still in the rest of the person's life going to school having your friendships grow your family life all of these things could be suffering so they're just finding local pleasure which we can assume the computer game might continually give the person by presenting them with this constant challenge in the end wouldn't be conducive to flourishing for that person because all of the other aspects of their life that are also valuable and important to them would end up being neglected so thinking about that from a dynamical or a computational it seems what's important is a kind of psychological flexibility or openness you're able to be poised between exploiting and exploring so the person that's playing the computer game is really just stuck in a kind of exploitative mode where that's continually something rewarding for them because they're continually challenging the challenges that they can solve but at the same time they're not exploring, they're not curious so there's not that kind of psychological flexibility which we think is required for eudaimonia or really living a good life with flourishing as a person so we think what you get out of active inference then is a neuro computational framework, a biological story about what it might be to as a person live a good life what might be required is this kind of openness and flexibility two other quick things on the tail of that one is there's a very old debate about higher and lower hedonia, the low hedonia is the sort of animal life that causes all the problems, higher hedonia but there is higher hedonia that leads you to whatever the good life is, that's kind of interesting from this perspective as well because you can think about it I think maybe as lower hedonia is aerodynamics in niches that leads you away from this sort of metastable attunement right so drugs of addiction is a good example of lower hedonia because it's good prediction error slopes but they're only local at the expense of your more integrated holistic governing state right but what would a higher hedonia be well it would be a local good slope that leads you to a greater integration so at the tail end of our paper and I didn't mention it here because it's a bit of a further topic but we suggested maybe thinking about zero sum and non-zero sum activities and their role in producing or restricting these sorts of metastable attunements and I think that's a good example where you can get hot about some small thing that reduces your dynamicism so you get this global rising in air even though you're locally succeeding I think it's a nice way to think about lower hedonia, higher hedonia is local good slopes that encourage and expand and generate a richer and more complex ecosystem of good slopes in your life so for instance think about the things that are endlessly rewarding love relationships you're never going to get to the end of being a good husband or wife you're never going to get to the end of being a good older brother or being a good dad, good mom you're never going to get to the end of those that's an endless thing or patience, compassion, love sympathy, empathy you're never going to get to the end of those things service to the people and to your community service is such a good example here if you feel good, if you get good slopes in serving your community you're just going to keep creating more connections that are going to create more opportunities for new kinds of slopes to be caught and encouraged so I think it's perfect almost it's a local slope that leads to lots more other slopes and just one other quick thing and I won't harp on about it and I think also that a system that's set up to be to have this more global dynamic regime working correctly I think they'll also tend to have more subjective happiness and it would be good to like go through this and build a detailed computational model because this part is just a little bit speculative but it seems to me that if you have a really rich diverse ecosystem of good slopes then if you reach the end of some particular slope like let's say you lose the video game if the video game is the only thing you have no wonder we have a problem regulating our emotions if you see this on YouTube all the time mom or dad turns off the system and there's an explosion and they say this is the only thing I have so you're very rigid, you're very fragile with only one outcome but if you have lots of stuff going on in your life then if you lose a good slope here it's just a matter of task switching over to the next good slope and so I tend to think that people who have diversity of good slopes in their life of course they're also going to have consistent more pervasive good feelings but you know that's a bit speculative so I love the look. This local global nested slopes descent I think it really helps us understand active inference but also understands or helps us understand how to apply active inference because it doesn't have to be an active inference model about emotion but that's the one we're talking about but it could be some other kind of active inference model so kind of just two other concordances I guess uh Cezanne Mahaly's flow and flow psychology it's like you mentioned selecting the right kind of errors and so it's like there's there's being in the flow state of playing chess or of doing a skill and then there's like getting the rock out of your shoe it's like you're still in this sort of uncertainty regime but one of those is just delaying you on the journey and the other one kind of is the journey so that and then one other it's dead on it's dead on actually John Varvakey at the University of Toronto and I are planning a paper using this exact model to think about flow as cascading cascading insight so every time you reduce you hit that good slope and then you learn that's what the good slope is you're learning and when you get to the end of that learning what it does is it opens up a new good slope that you learn down which opens up a new good slope which you learn down and that happens in a sort of cascading a cascading pattern yeah I love that so that that perfectly comes to the second example the cascading insight like so many people whether they're in academia or just curious about something when they get a quote answer they say well you know just brought up so many more questions like now I know how much this planet is in diameter but now I'm curious about something else and so it's like that short term minimization is like disciplinary research like we had 25% efficiency on this chemistry extraction we got 7 it's a local who can who can say that's not a local improvement but then that global is the transdisciplinary understanding and as the research ecosystem if we just dove down local disciplinary rabbit holes we would end up with just a wreck but at the same time we do need to do some local gradient descent otherwise we're not extracting that compound so what can we build our transdisciplinary theory on and so it's just like very cool how we can have a positive I don't know can it be neutral and positive at the same time yeah yeah why not yeah it's positive in the sense that it's not negative I but I love that and I think that's a I think it's a role that philosophers are increasingly playing you know what you might call synthetic philosophers philosophers who are trained in various scientific disciplines the neighboring scientific disciplines around a certain topic like Dan Dennett you know who are able to zoom out and create good overall frameworks that sort of that bring together various disciplines as well as to the special sciences that zoom in and discover locally I think one of the interesting things is the three of us here are all philosophers so we will come at active influence through philosophy but through trying to do that synthetic kind of philosophy where we're integrating philosophy with neuroscience with biology and so that's a common theme and another thing that I think is interesting among the three of us is that we're all thinking about active influence is something that is environment involving so one of the crucial interesting things I think in his work on the psychotic mark of blankets is what happens when you're you're you have a free actually minimizing system that's insulated or isolated from and what's coming in from the environment and how that can induce pathologies a lot of the things we've been talking about today are about kind of designing an ecosystem where being in that kind of environment would help you to flourish as an individual so what would the how would you need social media to look if it was creating the kind of social pathologies that we're seeing today with the misinformation problems and all of the problems that are arising from polarization you had a social media environment in which people flourished so we think that these ideas about active influence if you have a philosophical perspective on them you can see how they could help you to think about questions maybe political questions as well so that synthetic perspective that philosophy can bring helps you to see how to take this framework and show how it can be applied in lots of different domains I think that was something that you touched on Daniel in one of your observations questions and I think is a valuable contribution that philosophers can make from using these ideas totally we see advances in philosophy alongside and deeply interwoven with governance and technology so here's just one last question I wanted to ask on that theme unless anyone has any other questions and this is to NS's section so you talked about thinking about using active influence as a model which was invoking this realism instrumentalism distinction that we had talked about in some other streams and I just wondered how will realism versus instrumentalism or potentially other philosophical ideas play out clinically like what will it look like to deploy something instrumentally and say that for example this is being supported by a mathematical framework but that's not what is per se happening inside of you it's just how we're instrumentally interacting with it I don't know just how will that kind of a philosophical nuance enter into the regular protocols of medicine or governance so I think that instrumentalism is very much aligned with dynamical systems to take on cognition or on brain activity so it doesn't need to be something that is highly mystical or that is like a framework that we use but then what can we do with it in medicine or in more experimental settings right it's just that it's not committed to more computational accounts of cognition or of the brain it's much more aligned with for example in this framework with the dynamical systems theory as in the case of being a tool that is useful to explain and describe an activity of scientific interest through some constructs that we know are constructs that we build for our own scientific practice and understanding so that's the idea it is useful in the sense that you can now put things together and make this activity that would otherwise be a blur intelligible in a scientific sense by using these constructs which you can think about it for example from dynamical systems theory Mark was mentioning these attractors for example which by the way I'd like to say something about that in the case of for example in psychopathology there is this kind of like fixed point or attractors that kind of like we need to change etc we are not saying that there is actually a fixed point or an attractor in the psychopathology but we are explaining this in this sense for our own or to make these patterns intelligible from a scientific point of view right so it's in this sense that you can use these instrumentalist accounts to say that you gain explanatory traction but you're not saying that actually there is a fixed point in the behavior itself does this help I think so I mean if someone looks at an economic model and it uses a regression it's not saying the economy is a regression but then there's something about the personal nature of the brain in our emotions where it feels almost like a model of emotion or emotional dynamics it's something that's been brought up many times I mean a lot of people like qualitative perspectives on these topics because it avert this whole discussion on whether the mathematical models are real or instrumental and it almost feels like now with that philosophical clarity to be able to frame the two different modes we have there we can distinguish very clearly instrumentalism from a sort of background of implicit realism and then that would maybe help speak to a broader set of people who have different beliefs and different structures for what they think is the underlying system Am I coming on this briefly? Yes. Yeah please. I think I take a different perspective for a minute on the realism, instrumentism issue so I start from the cybernetic idea of the good regulator theorem that Ashby put forward and I think that gives us a way of thinking about how values are actually built into living systems and so the idea of being a model of the environment so not just having a model but actually being a model and how you can be a good model so what does it take to be a good model? Well there has to be as much complexity in the model as there is in the environment that you're interacting so I think what we get out of Ashby's good regulated theorem is an idea about what it is to be well adapted to the environment and I don't take that just to be like a model of life or of systems that are well adapted but that we can actually get some principles that tell us how would a system need to be organised in order to be well adapted and those principles I mean they are a model I agree but they actually tell us something about the organisation that a complex adaptive system would need to have if it was to remain well adapted over time so if it was to be a good model and then we can get some kind of valuable normativity out of that because notice that there's a notion of goodness that we're thinking about what would it take for a model to be a good model of this environment for it to remain well adapted when there's a kind of value or normativity implied there as well so I think all of that is not just about finding explanatory models that are good job for us as scientists but it's also telling us something about what it is to be alive what it is to be a good cogniser so that we can begin to get some kind of values out of biology and I take that to be a kind of realist project because I think that we can get some notion of value out of free energy minimisation in fact I think the notion of free energy is itself a kind of value for a system that tells a system what it is to know to be well adapted it's well adapted if it is first of all minimising free energy over time but we saw from Mark's talk and from our paper that on its own it's sufficient you need in addition this kind of maintaining less stable poise so all of that is just to say well I think we can get some something more than just a model out of free energy principle we can get some principles which help us to make sense of how there could be values in a world of facts and it helps us with that question it may help us to understand what it is to live a good life so how could you have natural systems that flourish so we could get some kind of naturalistic account of ethical values so that would be my take on it and I think that's kind of different from instrumentalism I think you can also get out of the free energy principles there's a debate here isn't there and that's one of the things that philosophy can bring in there's different perspectives you can have on what the free energy principle means and then we can fight about amongst ourselves as philosophers great closing point we respect and value the content but then also there's that philosophical prior that the space for the debate and the space for different perspectives is itself important so with that thanks to all of you this was an awesome session you're always as well as anyone else welcome to come back share any other exciting research or questions so great times and I hope everyone watching live in the replay appreciate this great session so thanks again everyone and see you next time thanks everyone