I mentioned in my review of Anatoly Levenchuk’s “Systems Thinking 2020” having some subsequent dialogue about common ground in other areas of the Psybertron agenda. A significant overlap is the work of Karl Friston (Free Energy Principle / Markov Blankets / Emergent Organism / Active Inference) in my reading of Mark Solms, and in Levenchuk’s case, where he and Friston are both members of the advisory board of “The Active Inference Lab”.
[Small world in itself – and yet in the days since, the concept of “systems thinking” is everywhere, from politics and biology, to consciousness and metaphysics. This is not going away. It was xxxx noticed back in January? I’d slipped into systems language quite naturally into ongoing dialogues. And, I made quite a thing of “systems architecture” considerations when interpreting both Solms and McGilchrist (independent) work in terms of (say) anatomical and functional brain architecture.]
In the dialogue above, Levenchuk shared a paper appearing to cast doubt on Friston’s use of Markov Blankets – “The Emperor’s New Markov Blankets” – ENMB (2021) for short here. Full refs:
- Bruineberg, Jelle and Dolega, Krzysztof and Dewhurst, Joe and Baltieri, Manuel (2021) The Emperor’s New Markov Blankets. Behavioral and Brain Sciences 1-63. [Preprint] doi:10.1017/S0140525X21002351
- Bruineberg, Jelle and Dolega, Krzysztof and Dewhurst, Joe and Baltieri, Manuel (2020) The Emperor’s New Markov Blankets. PhilSciArchive [Preprint]
- Bruineberg, Jelle and Dolega, Krzysztof and Dewhurst, Joe and Baltieri, Manuel (2021) The Emperor’s New Markov Blankets. Behavioral and Brain Sciences 1-63. [Preprint] doi:10.1017/S0140525X21002351
(More dialogue below in the Post Notes.)
It’s a substantial paper, 48 pages with some pretty heavy maths as well as arguments of principle. In fact when I read the parts of Solms where, amongst other things, he used (Freudian) mathematical notation additionally developed with Friston, I noted that it was perfectly possible I wasn’t properly understanding Friston’s arguments. Whilst it chimed intuitively with my own understandings, I wasn’t well placed to say whether it was formally right, one way or another – an occupational hazard in this kind of multi-disciplinary research.
“I’ve also kept in [my reviews] lots of technical specifics which I probably don’t understand as Solms intended, primarily to allow me later checking against other resources” (Myself, earlier.)
Well here is an opportunity 🙂 to respond to the “ENMB Paper” quoted directly below:
“This web of formalisms (Free Energy Principle, Markov Blankets, Active Inference) is developing at an impressively fast pace and the constructs it describes are often assigned a slightly unconventional meaning whose full implications are not always obvious. While this might ironically explain some of its appeal, as it can seem to the layperson to be steeped in unassailable mathematical justification …”
“We will argue that although this approach might have interesting philosophical consequences, it is dependent upon additional metaphysical assumptions that are not themselves contained within the Markov blanket construct.”
“In our view the FEP literature consistently fails to clearly distinguish between the ‘map’ (a representation of reality) and the ‘territory’ (reality itself). This slippage becomes most apparent in their treatment of the concept of a Markov blanket.”
“… a broader tendency within the FEP literature, in which mathematical abstractions are treated as worldly entities with causal powers.”
“[Friston’s is] a new and largely independent theoretical construct that is more closely aligned with notions of sensorimotor loops and agent-environment boundaries.”
“Inference within a model, as opposed to inference with a model, seeks to understand inference as it is physically implemented in a system, and places literal Markov blankets at the boundary between the system and its environment. The ‘model’ within which these Markov blankets are used is usually understood ontologically: here the map is the territory – the system performing inference is itself a model of its environment, and its boundary is demarcated by Markov blankets.”
“This procedure of attributing to the territory (the dynamical system) what is a property of the map (the Bayesian network) is a clear example of the reification fallacy: treating something abstract as something concrete (without any further justification) … we propose to distinguish between ‘Pearl blankets’ to refer to the standard ‘epistemic’ use of Markov blankets and ‘Friston Blankets’ to refer to this new ‘metaphysical’ construct. While Pearl blankets are unambiguously part of the map (i.e., the graphical model), Friston blankets are best understood as parts of the territory (i.e., the system being studied).”
“As a general rule, one should not mistake the map described by a model for the territory it is describing: a model of the sun is not itself hot, a model of an organism is not itself alive, and so on.”
OK, so again, without going through any of the mathematical rigour – itself unassailable by me – an important issue is indeed covered by the extract above, that there are metaphysical (ontological) premises possibly unstated in the work of Friston (and Solms), that might “appear to break this general rule without any further justification”. However these are quite explicit here.
Solms’ own response is categorical without any further metaphysical justification.
I have read [the paper]. I don’t think the Markov blanket formalism is a map of a territory but a description of the causal dynamics that actually exist in a territory. The territory in question is the (monist) functional organization of both brain and mind.
The ‘territories’ are the observable mental and neural phenomena. What they are calling the ‘map’ is, for me, the underlying functional system that explains those phenomena. This explanatory level (the functional organization of the system) cannot be observed; it must be inferred.
It is a dualist position. The formalism describes the actually causal ontology. As Galileo said: the book of Nature is written in the language of mathematics.
(Solms in Twitter exchange.)
I don’t buy the Galilean / Platonic argument as definitive, but it reinforces that this is not an accidental error in this school of work, but a deliberate act that needs to be understood as such. Sure the “book” of nature may be written in maths, but maybe not “nature” itself?
Good question, from any small boy not seeing the emperor’s clothes.
But it’s not necessary to analyse exactly what Galileo, or Plato before him, was asserting in any specific detail. That general rule of not confusing the map with the territory – not falling for the reification fallacy – is good advice and indeed is ancient advice. A Buddhist might point out that “the finger pointing at the moon is not the moon”. In science generally, our models in mathematical constructs used to represent, analyse and predict data about the real world are contingent approximations to the behaviour of that real world, but they are not it. Physics isn’t the real world, it’s our best current model of it. In any number of more mundane engineering applications, especially those that get implemented in analytical and operational computing applications, we constantly have to remind ourselves that the model is only a model, not the real thing, however seductive the virtual reality might be.
Dennett (much cited here) is among those acknowledged as providing advice to the ENMB paper, without any specific reference and, given his views on disembodied information and computation – independent of any physical layer – in his own “evolved consciousness” story, I’d be interested to know his actual views on this argument. I’m pretty certain he uses very similar arguments to Friston and Solms, as I do too.
What we are saying is that in this model, the computation, the sensing of information and algorithmic processes of the systems and subsystem components, with and without Markov-blankets, is quite literally happening. These information entities and processes are more fundamental than the physical models which self-organise and emerge from them. This is indeed a metaphysical claim, whether or not explicitly stated as such by every user.
In these theories, in my own metaphysics as well as Solms says above, the information processing is the territory, the foundation of the territory itself not just a map of it. Though obviously like any model we have also plenty of other information representations used to describe and present (map) the model and its processes to human audiences.
Friston and Solms (and myself) are not unique here, as the ENMB paper acknowledges, there are many philosophers and cognitive scientists with information-and-computation-based ontologies of reality. Integrated Information Theory (IIT after G Tononi) is one well developed example, but these are part of a wider movement. One corollary of these foundational (metaphysical) information-based ontologies is that both the physical (body) and mental (mind) worlds and their causal relations are explained by the same underlying metaphysics. A credible monism where dualism has stubbornly continued to exist. (Also a lot of new interest in various versions of pan-psychism in the 21st C, and again these theories provide an information-based “pan-proto-psychism” that may better support these.)
In many ways it’s good that the ENMB paper exists, because it is ringing an important alarm bell that more people in both science and philosophy should wake-up to how radically important these not-so-new theories are.
Thanks for the warning ENMB, but what you are describing is exactly what we’re doing.
One source here on Psybertron that I’ve not really developed yet is John C. Doyle, a control systems guru I’ve mentioned being impressed with before – he’s written the text-books and is much cited in papers – but he hasn’t written for a generalist public so he’s quite low profile if it’s not your field. He’s very much a systems thinker looking for architectural abstractions yet using very real-life examples to illustrate. In this 30 minute talk – very dense / terse / rushed, packed with content easy to miss if you’re not concentrating – the last 15 minutes is very interesting. Very clearly joining up issues of multi-layer systems optimisation and evolution (Levenchuk) with human situational awareness and responses based around the visual field and the speed of saccade eye-movements (Solms)?
Anatoly Levenchuk Comments:
2. There are not only “map-territory” distinction and representation relation that may be confusing. There are functional object — physical object distinction with implementation/realization relation. And you should decide: what type of relation each of authors mentions.
IG: Ah, yes. Not always explicit in every discussion, but pretty fundamental that the systems / architecture view is functional – you maybe saw my comments on trying to get a Brain Atlas that held to this schematic view. See later comments on process-based relations.
3. A day ago I did post about phys-math-modeling and compactification/universalization of knowledge. It suggest more long chain of ontology modeling: physical object from domain that is classified/annotated by type of physical object (ToPO) from physics textbook and then this ToPO is represented (or classified/annotated, if you prefer it) by mathematics/abstract object. I am not mention about functional object option here, it is enough complicated with this. Mathematics is foundation and upper ontology, physics is middle ontology, domain objects is working ontology.
Thus you can parse phases like “As the locus of molecular, thermodynamic, and bioelectric exchange with the environment, the cell membrane implements a Markov Blanket (MB) that renders its interior сonditionally independent of its exterior (Pearl 1988; Clark 2017); this allows the cell to be described as a Bayesian active inference system (Friston 2010, 2013; see also Cooke 2020 for a variation on this approach)” — this “implements” means classification relation (but easily you can go along 3D extentionalism and try functional-physical object “implementation/realisation”).
Phrase I took from https://chrisfieldsresearch.com/min-phys-NC-2021.pdf
My text (sorry, in Russian) here: https://ailev.livejournal.com/1621997.html
IG: Excellent. The chain of causality, with emergent layers separated by Markov blankets is the model I’ve had in mind all the way through – even before I’d consciously heard of Markov blankets 😉 As you know my interest is going back to metaphysical foundations, but yes, even with a functional bias / preference we must get to the functional-physical realisation in the real world. Think I’ve come across Fields and Glazebrook before but yes … at root in my model, “(All) physical interaction is information exchange”. (That’s precisely why information is more fundamental metaphysically 🙂 )
4. “, the information processing is the territory, its foundation not the map of it, though obviously we have plenty of other information representations used to describe and present (map) the model and its processes to human audiences” — you refer here to “information processing” and I can point you to:
— “Integrating information in the brain’s EM field: the cemi field theory of consciousness”, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7507405/ — “I describe the conscious electromagnetic information (cemi) field theory which has proposed that consciousness is physically integrated, and causally active, information encoded in the brain’s global electromagnetic (EM) field. I here extend the theory to argue that consciousness implements algorithms in space, rather than time, within the brain’s EM field”.
— “Types as Processes, via Chu spaces”, We match up types and processes by putting values in correspondence with events, coproduct with (noninteracting) parallel composition, and tensor product with orthocurrence. We then bring types and processes into closer correspondence by broadening and unifying the semantics of both using Chu spaces and their transformational logic. Beyond this point the connection appears to break down; we pose the question of whether the failures of the correspondence are intrinsic or cultural. — https://www.sciencedirect.com/science/article/pii/S157106610580475X?via%3Dihub (и дальше по этой линии Information flow in context-dependent hierarchical Bayesian inference, https://chrisfieldsresearch.com/contextual-pre.pdf)
IG: Thanks for these. I’m sceptical of “CEMI” and I don’t consider it necessary – it’s an information field, whatever the physical substrate, EM or otherwise – but I may follow-up, since Chris Fields is someone I have time for. Thanks.
Hope I paid my debt of commenting your posts )))
Sorry, but I am not sure that tell you something substantional about Markov blanket. Your post show that you already understand difference between abstract MB и physical one, I can only add complexity of ontology choices with functional object variant )))
IG: Actually – I was hoping you’d comment on the two prior posts on Solms – but YES – we have effectively covered the same ground. I appreciate your confidence I’m understanding this – or that at least we’re both misunderstanding it the same way 😉 Many thanks for the dialogue.
In my view “process” (something that flow, functional diagrams always with some flow/current in it) is good heuristic that we deal with functional (run-time) objects, not physical/product/module (construction time) objects.
I often have talks with “only process” people (e.g. category theory or other “transformations first”). Common thinking is about objects (of attention!) and relationships (processes) and IMHO this is supported by wetware in a brain. To the thinking you need both, but with only objects it became metaphysical (especially if you have no at least 3 timescales — evolution, learning/adapting and run-time) but with processes only you have absence of attention anchors for wetware. Therefore both but things first, process second.
4D dimentionalism is good for integration of object vs. process false dichotomy. Process is related not only time and function, but also space and physical objects!
IG: Yes understand that 4D model and that in the real day-to-day world we interact with space and “things” – but I am (after Whitehead) being quite radical here – metaphysical again 😊. (These “things” only emerge from networks of “events” at the information level … longer story).
Yes, works of John Doyle is relevant here — https://scholar.google.com/citations?hl=en&user=C6DtGmMAAAAJ&view_op=list_works&sortby=pubdate
IG: In my more general public dialogues – as opposed to researching technical papers – I find very few who have heard or or understand Doyle’s multi-layer architectural-optimisation. Once we accept layers (Markov blankets) as REAL, then I think his view is VERY IMPORTANT to so much evolution / self-organisation of ALL systems. This is amazing convergence from the practical engineering level right back to fundamental physics – as information and computation 🙂 Many thanks again.
12 thoughts on “The Emperor’s New Markov-Blankets?”
This is difficult material for me, but in all of it the map-and-territory metaphor is perhaps perversely instructive. All human understanding is a map. The territory it surveys is ineffable. Attempts to describe a territory exhaustively by means of a map are doomed to failure. In recognition of this, the constructs and ontologies we impose must forever be flexible. Thus it may be true that everything is information, within the framework of a certain approach — but within the framework of another, equally useful approach, you still can’t stub your toe on it.
I agree with someone or other (can’t quite make out the back-and-forth) that the temptation to reify “information,” or for that matter anything we talk about, is a constant hazard.
Hi AJ – yes finding things difficult is the occupational hazard whilst trying to develop cross-discipline understanding, I wouldn’t worry too much.
You get the point anyway.
For ANY model/framework information is the map, a representation.
For this particular model/framework information is also the foundation of the territory itself.
And as you say each model/framework has its domain of usefulness.
And with any model/framework the “constant hazard” is to mistake the map for the territory and its parts – the reification fallacy was in the original ENMB paper.
THE KEY THING here is the information-based model is most useful, IF you need one that explains BOTH the objectively-physical and the subjectively-mental.
You can understand that too, even if the reasoning why is harder?
The claim that an information-based model explains both the physical and mental is easy enough to understand — as is the claim that a particle-based model is adequate to both purposes. Both are equally interesting, which is to say: not very. It’s all about the reasoning.
Unfamiliar material can be difficult, but I’m not embarrassed or troubled by that fact. I’m just trying to be clear about it, in order to place my contributions in the proper context.
To say that “information is the map” could be construed (or misconstrued) as similar to “the paper and ink are the map.” But now we have devolved to talking about the map of the map, as it were; that is, how one represents the map. Yet this is not to stray further from the territory; we are at exactly the same distance, only now we are considering the map as the territory.
To say that the information thus equated with the map is also the foundation of the territory which is mapped, is to confuse the information with what it is information about, much as we might confuse paper and ink, or lines and text, or two-dimensional relations between blobs of colour, with a map in its capacity as a function, a sort of verb. One is soon led to the position of suggesting that the information is not information about anything, but rather, that the information is information about what the information is about. This is indeed a difficult doctrine.
Nope, I think you are missing the key point:
You said “To say that the information thus equated with the map is also the foundation of the territory which is mapped, is to confuse the information with what it is information about …”
It might be confusing, but I can assure you we’re not confused.
The information ABOUT the territory is the MAP (paper ink pixels whatever, a representation)
(Ancient wisdom, not contentious.) We are not confused.
We are ALSO saying for this particular definition of the territory, information is ALSO the foundation. Different information obviously.
So easy to confuse WHICH information we might be talking about at any given time, but we are talking about two distinct instances.
One which IS the territory, which the territory comprises, and another which is about the territory, a descriptive representation.
The question of what this underlying information is about, is a good one – it may not strictly be about anything but itself. [In fact the further atomic we go (back to Euclid and Democritus) about is really just saying which relations … ABOUT is an emergent concept at higher levels of complexity … now we’re explaining / talking about what the metaphysics / ontology actually is … “the reasoning”]
That clarification helps, thanks. It wasn’t obvious that “information” was being used in different senses: one the one hand, as representing or describing something (information about), and on the other, as constituting something (information that forms or informs, as in information theory).
It’s the latter sense, I think, in which you talk about “underlying information.” But I remain skeptical of the scientific or explanatory usefulness of speaking as if it informs nothing other than itself. There are contemplative or revelatory uses for such a mode of thinking in shaking our core assumptions, particularly if they are materialistic: the sort one encounters in Buddhist ideas of nothingness. When we come to, say, an account of quantum phenomena, this mode can help break down inhibitions in our thinking. But it is nevertheless paradoxical, and for practical purposes one reaches for something that can be informed, other than the information itself.
That’s where I am at the moment, anyway. Meanwhile, Solm’s The Hidden Spring has arrived, and I’m a few chapters in. Thanks for bringing it to my attention. It’s very interesting, and so far quite clearly written. I’ve long been a fan of Damasio, having bought a copy of Descartes’ Error when it came out. I was led there by my interest in Kierkegaard’s emphasis on subjectivity. It’s encouraging to see the return of phenomenalist or psychological influences in the attempt to understand our reality.
You concluded the post with, “Thanks for the warning ENMB, but what you are describing is exactly what we’re doing.”
I haven’t finished Solm’s book, but having read the chapter about Markov blankets, I felt emboldened to tackle the ENMB paper. (I notice that the linked PDF was to a 2020 version. Apparently there is a more recent version, less technical and more philosophical, but I don’t have access. That’s a pity, because of course it’s the philosophical part that’s of interest.)
The point of the paper, I thought, was to argue that Friston’s work with Markov blankets misconstrues them in a way that smuggles in metaphysical assumptions. Based on this work, Solms purports to show that “self-organization brings participant observers into being.” The question is whether the emergence of “participant observers” has really been demonstrated, or simply smuggled in as a hidden assumption involving the misunderstanding of “Pearl blankets” (a mathematical convenience for simplifying calculations involving many variables) as “Friston blankets” (a portrait of actual “things” as defined by informational boundaries). Here, what counts as a “thing” is unfortunately subject to the usual ontological ambiguities. As the paper points out, the selection of the “internal” node is arbitrary, and even within a single simple example network, cannot be cleanly delimited.
If what you’re doing is smuggling in metaphysical assumptions to beg the conclusion, then I can’t possibly comment. But I don’t think that’s what you intend, and so I’m not sure you got the point of the ENMB paper.
I’m certainly missing your point here? You seem to be saying things about the paper I would agree with.
Not sure about “begging” the conclusion (in the technical philosophical sense).
The “decision” to treat Markov-blankets (and all the algorithmic mathematical computation processes) as delineating real things in our ontology is the metaphysical presumption? (And that’s a good thing). They’re only “smuggled in” in the sense that’s not explicitly stated (by Solms)
“Begging the conclusion” was my unfortunate mangling, for syntactic convenience, of “begging the question,” but I do mean the technical sense of a logical fallacy, where the premisses are assumed in the conclusion. The metaphysical assumption is that we are dealing with distinct ontological entities with distinct boundaries. Whether we call them “physical” or “functional,” the idea that they are blanketed, or fully enclosed, is not true to the original conception of Markov blankets, which involve for the outputs (in this case, the “active” nodes) additional influences called “co-parents.”
Incorporating co-parents into Solms’ model could be a promising direction for a more complex understanding of the interaction of what Whitehead called occasions, allowing for a multi-dimensional aspect. As it is, the concept of co-parents is omitted completely, allowing the model to posit a self-contained, independently defined entity, a”Friston blanket.” The metaphysicial assumption involves the supposition of such an entity, in the first place; the selection of nodes to be included; and the designation of the locus of interest that forms the centre of the collection. In a network with co-parents, these choices would be somewhat arbitrary. The ENBM paper is concerned with the misapplication of true Markov blankets as a way of lending authority to the metaphysical assumptions — the arbitrary ontology — afforded by the conceptual omission of co-parents. That’s how I read it, anyway.
I’ve just finished the penultimate chapter of The Hidden Spring, on “The Hard Problem.” Solms seems to hope that moving the discussion from the cortex to the midbrain, and from cognition to feeling, will somehow help with the hard problem. Meanwhile his argument seems to straddle a desire for physicalist reductionism (see the end of Ch. 6) with an appeal to psycho-physical parallelism, in an uncomfortable way. While he seems to think the hard problem is not so hard, he certainly has a hard time addressing it, causing him to point out many times in the course of the book that it still remains to be explained (away).
In the final analysis he offers a mere gainsay: “The biological function of feelings like hunger is nothing mysterious; and their somthing-it-is-likeness is not especially difficult to explain. Just follow the logic of the free-energy minimisation where it leads for self-organizing systems like us. Given our multiple needs, complex and perilous environments, wide choice of possible actions and ability to perform only one or two of them at any given time, we should expect to have an inner world, built for the purposes of deliberation and choice.” (p. 267) We do have one, and he appeals to this when he asks us to “cross the Rubicon” on p. 210. But why should we expect experience from mechanism?
Eventually I hope to write a review or critique for my blog.
“The ENBM paper is concerned with the misapplication of true Markov blankets as a way of lending authority to the metaphysical assumptions — the arbitrary ontology — afforded by the conceptual omission of co-parents. That’s how I read it, anyway.”
Yes, but all ontologies are arbitrary – what we “deem” to exist based on some presumptions. It is quite correct, we might use the Markov-blanket maths to explain the mechanisms at work in our ontology, but our metaphysical assumptions are distinct from “look Markov-blankets!”. What those metaphysical assumptions are does need saying, even if Solms doesn’t address this – unsurprising since he’s not doing metaphysics? I’m “glad” the ENMB paper pointed this out.
All his stuff about “the hard problem” and Chalmers generally, is for the birds – irrelevant. I’m guessing his publisher or early reviewers probably advised he’d have to do this to be taken credibly. But as I said, he dissolved any “hard problem” en-passant. Our consciousness IS our subjective view. (Excluded from science by orthodox scientists unwilling to cross that Rubicon – the whole issue.)
I think Solms still has one foot in that scientific orthodoxy when he says at the end of Chapter 6, “We scientists typically do so by turning not to metaphysics but to physics. The answers we require are to be found in the physics of entropy.” This is presumably meant to preface his entirely physical and fully quantifiable account, equations and all.
Nevertheless, to complete his argument, we are to cross the Rubicon and “empathize” with a system which is thereby assumed to have experience. “This leap is justified” by the two facts that “experience can only be registered from the subjective perspective” and that “I have shown, in formal and mechanistic terms, how the subjectivity of self-evidencing comes into being.” (p. 211) Well, no; all he has shown is how the alleged subjectivity might come into being, if we grant first that there is subjectivity (by acceding to his plea for empathy). But we are still free not to cross that Rubicon, and to rest with his fully quantifiable, entirely physical account sans subjectivity. The hard problem is why we need to cross the bridge at all. For most, an appeal to empathy is not compatible with “We scientists. . . turning not to metaphysics but to physics”. If there were equations for our empathy, I missed them. I can’t make out which side of the bridge he wants to be on, but it can’t be both.
I think he is signalling the opposite with that “typically”. He needs his scientist colleagues kept on side so long as possible. If he tells them at the start he’s planning to go “off piste” from scientific orthodoxy none will follow him until they get to that plea for “empathy”.
He can only stick “sans subjectivity” IF the necessary metaphysics can be justified without it – and it can’t.
Rather than “might” I say “can” (as in possible) – then I adopt physics as a set of counterfactual rules – things that “can” happen.
(Sorry a bit rushed this evening …)
Actually what he has shown is how subjectivity might be present, not necessarily how it might have come into being. If in order to grant experience we rely on empathy, then the appearance or emergence of experience in connection with a system is not even a question, since the notion of a causal, or for that matter intrinsic, relationship is technically irrelevant to its assumed presence.
There’s no hurry to this back-and-forth. Certainly I haven’t always responded promptly, what with other things getting in the way. I also find it’s helpful to take time to digest and ponder before committing to a response.
I’ve worked up some initial thoughts for my blog, using my weirdly oblique reviewing style, and they should be posted by sometime tomorrow.