That bloke

Is Karl Friston that bloke? You know who I mean. That really clever bloke, master of some academic field. He’s used to understanding a few important things far better than the rest of us. But when people in the pub start discussing philosophy of mind, he feels wrong-footed in this ill-defined territory.

“You see,” he says, “the philosophers make such a meal of this, with all their vague,  mystical talk of intentions and experiences. But I’m a simple man, and I can’t help looking at it as a scientist.  To me it just seems obvious that it’s basically nothing more than a straightforward matter of polyvalent symmetry relations between vectors in high-order Minkowski topologies, isn’t it?

Aha! Now you have to meet him on his own turf and defeat him there: otherwise you can’t prove he hasn’t solved the problem of consciousness!

I’m sure that isn’t really Friston at all; but his piece in Aeon, perhaps due to its unavoidable brevity, seems to invoke some relatively esoteric apparatus while ultimately saying things that turn out to be either madly optimistic or relatively banal. Let’s quickly canter through it.  Friston starts by complaining of philosophers and cognitive scientists who insist that the mind is a thing.   “As a physicist and psychiatrist,” he says ” I find it difficult to engage with conversations about consciousness.”  In physics, he says, it’s dangerous to assume that that things ‘exist’; the real question is what processes give rise to the notion that something exists. (An unexpected view: physics, then, seeks to explain the notions in our minds, not external reality? Poor old Isaac Newton wasn’t really doing proper physics at all.) Friston, instead, wants to brush all that ‘thing’ nonsense aside and argue that consciousness, in reality, is a natural process.

I’ve spent some time trying to think of anyone who would deny that, and really I’ve come up empty. Panpsychists probably think that at the most fundamental level consciousness can be a very simple form of awareness, too simple to go through complex changes: but even they, I think, would not deny that human consciousness is a natural process. Perhaps all Friston means is that he doesn’t want to spend any time on definitions of consciousness; that territory is certainly a treacherous swamp where people get lost, although setting out to understand something you haven’t defined can be kinda bold, too.

To illustrate the idea of consciousness as a process, Friston (inexplicably, to me anyway: this is one of the places where I feel something might have got lost in editing) suggests we swap the word and talk about whether evolution is a process. Scientifically, he says, we know that evolution isn’t for anything – it’s just a process that happens. Since consciousness is a product of evolution, it isn’t for anything either. I don’t know about that; it’s true it can’t borrow its purpose from evolution if evolution doesn’t have one; the thing is, putting aside all the difficult issues of ultimate purpose and the nature of teleology, there is a well-established approach within evolutionary theory of asking what, say, horns or eyes are for (defeating rivals, seeing food, etc). This is just a handy way to ask about the survival value of particular adaptations. So within evolution things like consciousness can be for something in a relatively clear sense without evolution itself having to be for anything. Actually Friston understands this perfectly well; he immediately goes on to speak approvingly of Dennett’s free-floating rationales, just the kind of intra-evolutionary purpose I mean. (He says Dennett has spent his career trying to understand the origin of the mind – what, is he one of those pesky guys who treat the mind as a thing?)

Anyway, now we’re getting nearer to the real point; inference.  Inference, claims Friston, is quite close to a theory of everything (maybe, but so is ‘stuff happens’). First, though, it seems we need to talk about complex systems, and we’re going to do it by talking about their state spaces. I wish we weren’t. Really, this business of state spaces is like a fashion – or a disease – sweeping through certain parts of the intellectual community. Is there an emerging belief, comparable to the doctrine of the Universality of Computation, that everything must be usefully capable of analysis by state space? I might be completely up the creek , but it seems to me it’s all too easy to use  hypothetical state spaces to give yourself a false assurance that something is tractable when it really isn’t. Of course proper defined state spaces are perfectly legitimate in certain cases. Friston mentions quantum systems; yes, but elementary particles have a relatively small number of properties that provide a manageable number of regular axes from which our space can be constructed. At least you’ve got to be able to say what variables you’re mapping in your space, haven’t you? Here it seems Friston wants to talk about, for example, the space of electrochemical states of the brain. I mean, he’s a professor of neurology, so he knows whereof he speaks – but what on earth are the unimaginable axes that define that one, using cleanly separated independent variables? He’s very hand-wavy about this; he half-suggests we might be dealing with a state space detailing the position of every particle, but surely that’s radically below the level of description we need even for neurology, never mind inferences about the world. It’s highly likely that mental states are uniquely resistant to simple analysis; in the state space of human thoughts every one of the infinite possible thoughts may be adjacent to every other along one of the infinite number of salient axes. I doubt whether God could construct that state space meaningfully, not even even if we gave Him a pencil and paper.

Anyway, Friston wants us to consider a Lyapunov function applied to some suitable state space of – I think – a complex system such as one of us, ie a living organism. He describes the function in general terms and how it governs movement through the space, although without too much detail. In fact after a bit of a whistle stop tour of attractors and homeostasis all he seems to want from the discussion is the point that adaptive behaviour can be read as embodying inferences about the world the organism inhabits. We could get there just by talking about hibernation or bees building hives, so the unkind suspicion crossed my mind that he brings Lyapunov into it partly in order to have a scary dead Russian on the team. I’m sure that’s unfair, and most likely there is illuminating detail about the Lyapunov function that didn’t survive into this account because of limitations on space (or probable reader comprehension). In any case it seems that all we really need to take away is that this function in complex systems like us can be interpreted as making inferences about the future, or minimising ‘surprise’.

It’s important to be clear here that we’re not actually talking literally about actual surprise or about actual inference, the process of some conscious entity inferring something. We’re using the word metaphorically to talk about the free-floating explanations, in a Dennettian sense, that complex, self-maintaining systems implicitly display. By acting in ways that keep them going, such systems can sort of be said to embody guesses about the future. Friston says these sorts of behaviour in complex systems ‘effectively’ make inferences about the world; this is one of those cases where we should remember that ‘effectively’ should strictly be read as ‘don’t’. It’s absolutely OK to talk in these metaphorical terms – I’d go so far as to say it’s almost unavoidable – but to talk of metaphorical inference in an account of consciousness, where we’re trying to explain the real thing, raises the obvious risk of losing track and kidding ourselves that we’ve solved the problem of cognition when all we’ve done is invoke a metaphor. So let’s call this kind of metaphorical inference ‘minference’. If we want to claim later that inference is minference, or that we can build inference out of minference, well and good: but we’ll be sure of noticing what we’re doing.

So complex, self-sustaining systems like us do minference; but that tells us nothing about consciousness because it’s true of all such systems. It’s true of plants and bacteria just as much as it’s true of us, and it’s even true of some non-living systems. Of course, says Friston, but for similar reasons consciousness must also be a process of inference (minference).  It’s just the (m)inference done by the brain. That’s fine, but it just tells us consciousness must tend to produce behaviour that helps us stay alive, without telling us anything at all about the distinctive nature of the process; digestion is also a process that does minference, isn’t it? But we don’t usually attribute consciousness to the gut. Brain minference is entirely different to conscious explicit inferences.

Friston does realise that he needs to explain the difference between conscious and non-conscious minferring creatures, but I don’t think he’s clear enough that the earlier talk of (m)inference is no real use to him. He suggests that in order to infer (really infer) the consequences of its actions, a creature needs an internal model. This seems quite problematic to me, though I’m led to believe he has a more extensive argument for it which doesn’t appear here. While we may use models for some purposes, I honestly don’t see that inference requires one (in fact, building a model and then making your inferences about that would be asking for trouble). I plan to go and catch a train in a minute, having inferred that there will be one at the station; does that mean I must have a small train set or a miniature timetable simulated in my brain? Nope. Friston wants to say that the difference between conscious and unconscious behaviour is that the former derives from a ‘thick’ model of time, which here seems to mean no more than one that takes account of a relatively extended period. The idea that the duration is crucial makes no great sense to me: the behaviour patterns of ants reflect hundreds of thousands or even millions of years of minference: my conscious decisions may be the work of a moment; but I think in the end what Friston means to say is that conscious thought detaches us from the immediate moment by modelling the world and so allows us to entertain plans motivated by long-term considerations. That’s fine, but it has nothing much to do with the state spaces, attractors and Lyapunov functions discussed earlier; it looks as if we can junk all that and just start afresh with the claim about consciousness being a matter of a model that helps us plan. And once that idea is shorn of all the earlier apparatus it becomes clear that it’s not an especially new insight. In fact, it’s pretty much the sort of thing a lot of those pesky mind-as-thing fellows have been saying all along.

Alas, it’s worse than that. Because of the confusion between inference and minference Friston seems to be saddled with the idea that actual consciousness is about minimising actual surprise. Is the animating purpose of human thought to avoid surprise? Do our explicit mental processes seek above all to attain a steady equilibrium, always thinking about the same few things in a tight circle and getting away from new ideas as quickly as possible? It doesn’t seem plausible to me.

Friston concludes with these two sentences.

There’s no real reason for minds to exist; they appear to do so simply because existence itself is the end-point of a process of reasoning. Consciousness, I’d contend, is nothing grander than inference about my future.

Frankly, I don’t understand the first one; the second is fine, but I’m left with the nagging feeling that I missed something more exciting along the way?

 

[The picture is Lyapunov, bu the way, not Karl Friston]

 

31 thoughts on “That bloke

  1. Hi Peter. I think “internal model” here is as per Conant & Ashby’s “internal model principle” eg

    http://folk.ntnu.no/skoge/prost/proceedings/ifac2008/data/papers/0568.pdf

    “The proper function of many biological systems requires that external perturbations be detected, allowing the system to adapt to these environmental changes. It is now well established that this dual detection and adaptation requires that the system has an internal model in the feedback loop.”

    Then mind is doing all the same things that life has always done, just a bit more knowingly.

  2. I plan to go and catch a train in a minute, having inferred that there will be one at the station; does that mean I must have a small train set or a miniature timetable simulated in my brain? Nope.

    In a sense, it kinda does, though: there must be some reliable mapping between whatever your mental content is, and what the state of the world at a particular time happens to be (i.e. such that a train arrives). Is there much more to ‘having a model’ of something than this?

    In fact, I think that most of our conscious, deliberate reasoning is essentially model-based: roughly, we understand something if we have a model of it—I understand, for instance, how to tie my shoes if I know a process which takes as input the untied state of my shoes, performs a series of manipulations on them, and ends with tied shoes. My internal model is such that states follow one another by an inference process, whereas in the real world, states are related by causality. (Note that I don’t mean here the sort of automatic, ‘unconscious’ shoe tying most of us probably engage in ordinarily, but the explicitly imagined process of how shoes are tied.)

    I think there’s something to the idea that many of the problems of consciousness arise because we conflate between the real world and our models thereof—and moreover, because our models are essentially computational entities, while the world isn’t. In a very rough sense, we feel we understand something if we can mentally ‘simulate’ it—that is, set up our mindstuff in such a way as to bear a particular relationship to an aspect of the world (or indeed, some abstract formal entity, such as the calculation of pi).

    But if this is right, and the models we use to think the world are computational, while the world isn’t, then it’s basically unavoidable that we bump into hard problems—issues where simply no model exists, or all models are necessarily incomplete in one way or the other. The simplest example for this is the generation of random numbers: there’s nothing to explain about it, no procedure to describe; some processes simply produce random numbers, and there’s no way to understand how. So we just live with the fact that a given system produced the digit 0, when it could just as well—for all our models tell us—have produced a 1.

    That’s I think the real problem with eliminativism: an eliminativist is somebody so convinced by their own models, that they conclude that since there is no model of how models are produced, there is actually no way to produce models.

  3. Thanks, David.

    My naive expectation is that a model independently embodies some of the features or behaviour of the thing modelled; I know some people have a looser idea. Of course, the looser your definition, the less you’re saying when you say consciousness has a model!

  4. I found a lot of individual points in Friston’s article to agree with, and it seems like, in a way, he was discussing the same consciousness as a prediction engine that Anil Seth and others have described. But I don’t think he did a good job tying it all together. The piece ultimately seemed somewhat muddled.

    I came away with the impression that he was trying too hard to reduce everything to a single explanation: inference. As someone who accepts the reductionist label (although not the eliminivist one), I think you can take the principle too far, to a point where you’re no longer providing an explanation. I think consciousness as a predictive simulation framework has a lot going for it, but trying to then reduce it further to some kind of inference principle seems like trying to explain the video game Angry Birds in terms of the physics of transistors. You can do it, but the result isn’t really a productive description of how the game functions.

  5. Almost forgot:

    “I’ve spent some time trying to think of anyone who would deny that, and really I’ve come up empty. Panpsychists probably…”

    My experience is that few scientists or philosophers of mind will explicitly admit they think of consciousness as a thing aside and apart from the physics, but much of their reasoning seems to have it as an implied assumption. Many panpsychists talk about consciousness as something similar to an electromagnetic charge or other fundamental force, a fundamental aspect of reality that is separate and apart from all the ones we currently know about. I’d have some sympathy with this idea if there were any evidence for it aside from people wanting it to be true. So, the notion Friston criticizes is, unfortunately, pretty prevalent in consciousness discussions.

  6. Let me swerve on a tangent- (as if the comments in fact were a bar room philosophical talk)- in Homo Deus, Harari seems confident in a similarly philosophically reductionist vein, that all brain processes boil down to algorithms.
    Is Friston from the same school as Harari? Harari stops short of denying consciousness or the real world. He seems a stop along that road

  7. Hi Peter,

    > I plan to go and catch a train in a minute, having inferred that there will be one at the station; does that mean I must have a small train set or a miniature timetable simulated in my brain? Nope.

    I rather think that it does.

    > My naive expectation is that a model independently embodies some of the features or behaviour of the thing modelled;

    Indeed, depending on what you mean by “embodies”. You certainly don’t have anything like a physical train in your head. But you have in your mind a representation of trains and of train timetables etc. That representation captures some of the features and behaviour of the thing modelled — i.e. everything you know about trains and train timetables. It’s not a physical model but an informational model of the kind computer systems or scientific theories use. I don’t think this is a particularly loose usage of the word ‘model’.

    Whenever you think of trains, I would say you are only doing so via this mental model, and so any intentions you have to things in the world ought to be understood as being rather indirect — all you are actually doing is manipulating your mental models, which happen to correspond closely enough to things in the world that you are able to catch your train on time, etc.

    That said I’m not terribly impressed with Friston’s argument overall, for much the same reasons as you.

  8. I think the idea that you need a model in your head to think about anything resembles the homunculus fallacy. It’s just that you’ve brought in a tiny version of the perceived instead of a tiny version of the perceiver.

  9. Jochen said:

    “I think there’s something to the idea that many of the problems of consciousness arise because we conflate between the real world and our models thereof”

    The opposite is true. The idea that there is at least one thing for which there is *no* difference between *model* and *reality* is the soluton to consciousness!

    If you go the comments in the ‘Aeon’ article, you’ll see I’m heavily involved , agressively and obnoxiously pushing my ‘time flow’ theory 😉

    I argue that there is a complete identity between *time flow* and *consciousness*. That is to say, I argue that time-flow literally *is* consciousness.

    The idea is that the conscious mind forms a *model* of time, and in the limit, there is literally no difference between the *model* of time and time itself. That’s because, your representation or model of time is itself in time. The model *is* the reality.

    The notion of consciousness as predictive inference is wonderfully simple:

    Imagine a chain with 3 links:

    (1) Bottom-up sensory data of past state
    (2) ?
    (3) Top-down abstract model of what you *expect* to happen

    The chain forms a model of past-present-future

    (1) > (2) > (3)

    The idea is that the missing link (2) is consciousness! It’s simply a model of the present moment , constructed by combining (1) and (3). Sensory data of the past is combined with your expectations about the future, consciousness is the result (model of present moment).

    The big twist I’m adding is that the model of time literally *is* time…consciousness is time itself!

    It’s as simple as 1,2,3

  10. I think the idea that you need a model in your head to think about anything resembles the homunculus fallacy.

    Indeed, it exactly is the homunculus fallacy, if you intend to use it to explain intentionality. Because all modeling is actually just the use of intentionality: you have some object, and a model system; between those two systems, you need some sort of mapping, taking states of the model to states of the object, in such a way that whenever a certain state shows up in the object system, then the image of that state under the mapping shows up in the model. That is, you take states of the model to stand for states of the system—but this is an instance of intentionality. Moreover, if you try to replace the mapping with something non-intentional, say, a computer, you just iterate the problem and incur an infinite regress: computation is just an instance of the modeling relation, obtaining between a physical system and a formally specified one.

    But that doesn’t mean you can’t make use of models in order to make sense of the world. If we for the moment take the fact that we can use our intentional faculty, whatever that may be, to map certain systems to others—say, pixels on a screen to arguments about consciousness—for granted, I think we can make headway on the question of how we formulate explanations of the things in the world: simply by coming up with a model that parallels them. There’s no homunculus problem here because we’re not saying that the mind creating a model is the explanation of how mental content can come to be about something out there in the world—which is just question-begging—, but rather, due to the capacity of mental content to be about things out there in the world, we’re able to create models of it. Models don’t explain aboutness; aboutness explains models.

  11. Jochen,

    If consciousness is a fundamental property of reality, present to some degree in everything (panpsychism), which is my position, it’s not a case of ‘models explain aboutness’ (I agree they don’t), but *nor* it is a case of ‘aboutness explains models’.

    Rather, it is the case that there is an identity relation between (a type) of model and reality, such that they are inseparable (in a limit).

    And I’m saying that the model/aboutness relation that explains consciousness is time-flow (the arrow of time, an aspect of time itself).

  12. Hi Peter,

    > I think the idea that you need a model in your head to think about anything resembles the homunculus fallacy.

    Perhaps it resembles it superficially, but it isn’t a fallacy, because there is no infinite regress once the little copy is of the perceived rather than the perceiver. I don’t need to postulate a model of the model or anything of the sort. The model you hold in your mind is just a representation of what you know about the world. There has to be such a model or else you could never be mistaken and you would know everything. There is only a model of the model if you choose to think about the model itself (as I am doing now). There is no infinite regress because eventually you get fed up of thinking about models of models of models of models of …

    I’m a computationalist (I know you are not, but you clearly know where computationalists are coming from), so I’ll make an analogy to a robot. Robots cannot navigate their environment (especially not in any sort of sophisticated way) without having a model of that environment. A self-driving car might for instance have access to digital maps and GPS data, which are a model of the roads and of the car’s location. I think this is the only way to have an intelligent agent interacting successfully in the world, and so I think human beings are no different.

    Hi Jochen,

    > Indeed, it exactly is the homunculus fallacy, if you intend to use it to explain intentionality

    I’m not necessarily trying to explain intentionality, although we have discussed that in the past so you know I think I can (and I know that you find my arguments entirely unconvincing). My point here is just to defend the idea that we need to have a model of the world in order to navigate it.

  13. My point here is just to defend the idea that we need to have a model of the world in order to navigate it.

    I don’t think that’s true: a stone rolling down a hill has no model of that hill, yet it manages to navigate it successfully—simply due to the causal influences it is subject to. Animals, at least at sufficiently low degrees of complexity (say, amoeba-level to be safe), also don’t have a model: chemotaxis just occurs due to concentrations of chemicals having a certain effect on the amoeba, ultimately causing it to alter its trajectory in a way that’s not essentially different from the stone changing course due to hitting a rock edge. Indeed, I don’t think there’s an intrinsic bound on the complexity of the behaviour possible using such a ‘model-less’ approach, and it’s entirely conceivable that we could navigate the world in much the same way we do, without any sort of model whatever.

    However, we do need a model to reason about the world, to formulate explanations for the behaviour of a system to ourselves: the system transitioned from state S_a to S_b, just as the model transitions from M_a to M_b, since M_a implies M_b must be the next state—the stone reaches the bottom of the ravine, just as in our model of the gravitational interaction, the state corresponding to the least potential energy is the ultimate endpoint of a system’s evolution. That’s, one might be tempted to say, why things fall.

    And for this sort of modeling, you need intentionality: you need to map S_a to M_a in such a way that M_a stands for (in the way of a symbol) S_a. There’s a sort of encoding/decoding relationship between the states of the model and the states of the system, and that’s what imbues the model with semantics. And it’s here that you indeed run into the homunculus fallacy upon trying to get rid of this mapping, or replace it by something non-semantical, like a computer: a computer is just a physical system such that there is some encoding/decoding relationship between it and some formally specified structure.

    Hence, if you want to use a computer in order to implement the encoding/decoding between model and system, you incur another encoding/decoding relationship; trying to get rid of that by means of another computer just kicks the can down the road, ad infinitum. That’s why there is no computational theory of mind: not because the mapping between a physical system and a formal structure is arbitrary, as Putnam, Searle, Bishop, and I during some brief discussions (;)) have maintained, but because there is such a mapping at all, which itself isn’t reducible to computation. That mapping might be one-to-one, but its presence—its necessity—still dooms a computational account.

    —————————–

    Zarzuelazen, I’m sorry if I appear to ignore you; but the truth is, I simply can’t make heads or tails of whatever you’re trying to propose, so I’m afraid I can’t really intelligibly comment on it.

  14. Hi Jochen,

    > I don’t think that’s true: a stone rolling down a hill has no model of that hill, yet it manages to navigate it successfully—simply due to the causal influences it is subject to.

    What I said was that I think one needs to have a model in order to navigate the world, especially in a sophisticated way. So let me expand on what I mean by that latter bit.

    To the extent that one navigates the world in a sophisticated way, one needs to have a model. The less sophisticated your navigation, the more stretched becomes an interpretation of your having a model. At the extreme, I would have to say that the rock “navigating” down the hill “represents” the direction of its travel with its momentum and “represents” the direction it is “trying” to go with the gravitational attraction between the earth and itself. This, I admit, is ridiculous and unhelpful. But I would say it is also ridiculous and unhelpful to describe the rock as “navigating” its environment. I wouldn’t describe it as navigating. It’s just passively rolling down a hill.

    An amoeba is a little more sophisticated than a rock. I agree it’s on the same spectrum, but the analogy to navigation and so a description of it as utilising a model is less stretched, though perhaps not by much. I don’t know a whole lot about the mechanisms of chemotaxis, but I would hazard to suggest that perhaps the concentrations of chemicals distributed around the interior of the amoeba can be interpreted as constituting a model of its immediate environment.

    But again, if it’s simply attracted towards a chemical or repelled from another, it’s a stretch to call that navigation, and to the extent that it is a stretch, it is also a stretch to say it has a model. But I would say that increasing sophistication of behaviour (and perhaps especially where behaviour is not only reacting to the immediate stimulus of the moment but also guided by “memory” of past stimuli and “anticipation” of future stimuli) necessarily makes it less and less of a stretch to describe the agent as having a model.

  15. There is no infinite regress

    I suppose the argument is that the model merely defers the job of apprehending the world. You need a model to understand anything, yet the model poses the same challenge as the the world and you need another model to understand that. And so on… (Can I say at this stage that I do accept models are used for some mental purposes.)

    There has to be such a model or else you could never be mistaken

    You see, I think this is a classic mistake which has led clever people astray for centuries. Typically the argument might go: when we experience a mirage of an oasis we can’t be seeing the oasis, because it doesn’t exist. So we must be ‘seeing’ something in our brain. In that case all we ever see is the proxy delivered by our brain, so we never see reality directly (if we did we’d be immune from error!) The point is that the mirage isn’t wrong; what’s wrong is your belief that it isn’t a mirage but an oasis. Error arises after perception, in our beliefs, not within it. We don’t even get to the stage where something is capable of being true or false until we get to the stage of beliefs about it.

  16. “say, amoeba-level to be safe”: the linked paper in the first post argues that bacteria have an internal model.

  17. Hi Peter,

    > I suppose the argument is that the model merely defers the job of apprehending the world. You need a model to understand anything, yet the model poses the same challenge as the the world and you need another model to understand that.

    That’s a plausible-sounding reconstruction of the argument, indeed. However I don’t agree that there is infinite regress.

    I would however go along with what you said to a large extent. If I want to understand X, I need a model of X. If I want to understand the model of X (i.e. as a model of X), then I need a model of the model of X.

    But I typically don’t need to understand the model of X, I only need to understand X. I do so via the model of X, but I don’t necessarily understand that I am doing so via the model of X. Subjectively, when I manipulate the model of X it feels like I am thinking only of X.

    I can work with the model of X without understanding it as a model of X because the model of X is just part of my mind. It is built out of whatever it is that minds are built out of (e.g. neural connections etc). Similarly, a self-driving car does not necessarily understand that its map data is a digital representation of a real physical road network outside it. It sees no distinction between its representations and the physical world. As far as it is concerned, it is simply “thinking” about roads as it finds a route by manipulating its map data — not that it actually has any subjective experience of what it is doing.

    > We don’t even get to the stage where something is capable of being true or false until we get to the stage of beliefs about it.

    I would say that a model comprises a number of beliefs about something. Your belief that the mirage is a nearby oasis is a part of the inaccurate model you have of your surroundings.

  18. Hi all,
    Some thoughts on this discussion.

    There’s nothing wrong with a homunculus provided each successively nested homunculus is less sophisticated. The frontal lobes are sometimes referred to as the brain of the brain, but they don’t have their own emotions or perceptions, depending on the other lobes and sub-cortical circuits for those. It just helps in reacting to those emotions and perceptions.

    My understanding of perceptions is that they are beliefs, just primal ones, our understanding of the sensory signals coming in.

    On modeling, the nesting isn’t infinite because eventually it ends with a mental reflex. All the modeling and simulations are adaptive extensions associated with distance senses (sight, hearing, smell) to increase the scope (in time and space) of what the mental reflexes are reacting to. (In pre-conscious species with nervous systems, such as primitive worms, the reflexes are all there is.)

    I think a lot of seemingly intractable issues with the mind become clearer if we widen our scope of investigation to include evolutionary biology.

  19. Disagreeable Me in 17:

    “I can work with the model of X without understanding it as a model of X because the model of X is just part of my mind. It is built out of whatever it is that minds are built out of (e.g. neural connections etc). Similarly, a self-driving car does not necessarily understand that its map data is a digital representation of a real physical road network outside it. It sees no distinction between its representations and the physical world. As far as it is concerned, it is simply ‘thinking’ about roads as it finds a route by manipulating its map data — not that it actually has any subjective experience of what it is doing.”

    We can think of conscious experience as the qualitatively rendered, subjective model in terms of which the world appears to each of us individually. All of what we know about the world, even the most abstract conceptual knowledge, is delivered via conscious experience, and the concrete presence of physical objects including the body is entirely a sensory-perceptual construction. Most of the time, like the self-driving car, we just take the model to be the world: the world appears to be directly presented to us. But we know from hallucinations, illusions (like a mirage), dreams and various sensory deficits that conscious experience is the epistemic intermediary between us as subjects who represent the world and the world as it is in itself, unrepresented.

    As conscious creatures, we always encounter the world fundamentally in terms of qualities which, curiously enough, aren’t found in scientific descriptions of it, for instance of how the brain does its representational work.

  20. You only need 3 levels of recursion to reach full modeling power:

    (1) Models of the world
    (2) Models of the mind
    (3) Models of minds modeling the world

    Infinite regress is avoided, because all further levels simply collapse to level (3) .

    Level 3 is ‘the theory of eveything’ – a conceptual framework that gives you a representation of all explanatory knowledge.

    Here’s the result (model of all knowledge):
    http://www.zarzuelazen.com/CoreKnowledgeDomains2.html

    Incidentally, this should tell you that cognition (mind) is a fundamental property of reality, since it enters into the conceptual structure of knowledge in an essential way. A physics theory alone is no-where near a true TOE.

  21. DM #14:

    The less sophisticated your navigation, the more stretched becomes an interpretation of your having a model. […] [T]he analogy to navigation and so a description of it as utilising a model is less stretched. […] [T]he concentrations of chemicals distributed around the interior of the amoeba can be interpreted as constituting a model of its immediate environment. [emphasis added]

    I think this sort of talk should raise several red flags immediately. Of course, any variable reliably correlating with some aspect of the environment can be interpreted as a model of that environment; but that tells us nothing at all about whether the system actually does possess any sort of modeling capacity. All it really says is that you possess the ability to model the system as containing a model of the environment; but all the modeling done here is just yours. Compare how Aristotle could interpret physical systems and processes as striving to achieve their natural place—but that doesn’t mean that physical systems really do any striving, have goals, affinities, and whatnot.

    No, what you need is a system such that it possesses a model regardless of any interpretation you might imbue upon it—otherwise, we just end in an infinite regress of interpretations (since after all, I can both interpret you as having a model of a system having a model, or as just producing bits of text). The system must model the world to itself, not merely be able to sustain an interpretation of possessing a model—the latter is trivial.

    Also, of course, with the inclusion of ‘sophistication’, you veer dangerously close to simply making an appeal to complexity: ‘yes, it’s clear that in every case that we can actually exhaustively analyze, there isn’t any modeling going on; but what if we mush enough of that together, such that we can’t really say with certainty what happens anymore?’. And well: perhaps. Maybe if you pile enough stuff together, quantitative additions will eventually produce a qualitative change—but if so, that’s really something that would have to be argued for, not merely assumed because we don’t know otherwise.

    But I would say it is also ridiculous and unhelpful to describe the rock as “navigating” its environment. I wouldn’t describe it as navigating. It’s just passively rolling down a hill.

    The thing is, though, that you can describe every physical system exactly in the same way as the rock—as a point ‘rolling down’ a landscape; it’s just that for more complex systems (with more than three degrees of freedom) that landscape is generally very high-dimensional. Still, all our navigating the world, sophisticated or not, just boils down to a ‘rock’ rolling down a very high-dimensional ‘hill’. Why would additional dimensions introduce meaningful navigation?

    (And before you point out that this is just an abstraction, some claim that it’s in fact more fundamental than our experience within three-dimensional space—for instance, the quantum wave function is defined in this higher-dimensional space, so anybody who’s a realist about that, has to accord it at least some measure of reality, as well. It’s just that, for some reason, all interactions depend on a parameter that can be most naturally framed as a three dimensional distance, hence, we can ‘order’ our experience by means of three dimensional space.)

    ——————————————-

    SelfAwarePatterns #18:

    There’s nothing wrong with a homunculus provided each successively nested homunculus is less sophisticated.

    I know that Dennett likes to make that argument (or liked, at least); but it’s never been really clear how this is supposed to work. After all, in the homunculus chain, you need the nth level homunculus in order to furnish a perception at the (n-1)st level; but if you want that chain to terminate, then you must have, for some n, the case that the nth level homunculus perceives nothing, yet the (n-1)st one does. If that works, no matter how dim the perception of this final homunculus is, then it seems that we don’t need any of it in the first place: just do what happens at this transition right from the start, and we’ll have perception—even if it is dim and degraded—without homunculi.

    And calling what happens at that level a ‘mental reflex’ really is just a collection of syllables—it doesn’t in itself tell us what, actually, happens such that this first spark of perception may bloom.

    Additionally, in the homunculus regress of the modeling relationship I proposed above, there really isn’t any measure of sophistication: at each level, you either have a mapping between states of the model and states of the object it models, or not; and if you don’t, the whole thing just doesn’t get off the ground.

    ——————————————-

    Zarzuealazen #20:

    You only need 3 levels of recursion to reach full modeling power:

    (1) Models of the world
    (2) Models of the mind
    (3) Models of minds modeling the world

    The problem here is that you’ve swept all the difficult stuff under the rug in the first step—how does one get to having ‘models of the world’? How do, in other words, mental states come to be about the world at all?

  22. Hi Jochen,

    > but that tells us nothing at all about whether the system actually does possess any sort of modeling capacity.

    I would push back on the suggestion that there is a clear objective fact of the matter on whether a system actually possesses a modelling capacity or not. Whether we describe what it is doing in terms of its using a model or not is a matter for our interpretation and not a matter of fact. When I assert that we have mental models I do so because this is the most sensible interpretation I think we can give of how we perceive and navigate the world.

  23. DM #22:

    When I assert that we have mental models I do so because this is the most sensible interpretation I think we can give of how we perceive and navigate the world.

    This reads contradictory to me. Surely, to describe/interpret something, one must have a model of it? But then, how can having a model depend on any particular interpretation? So to me, this seems to say ‘whether you have a model of yourself depends on your model of yourself’.

  24. Jochen #21,
    On the homunculus, it seems like you’re treating awareness as an irreducible thing. This is a common sentiment in consciousness discussions (see my first comment in this thread), but I think it isn’t backed up by neurological case studies. People can lose aspects of awareness, such as the ability to perceive motion, the feeling of being in their body, the ability to feel certain emotions, and many other aspects of consciousness. They can lose the ability to focus attention or to introspect effectively.

    Understanding awareness requires that we be willing to unpack the term “awareness”, to examine its components. And then be willing to unpack the resulting components, until we arrive at components without any functionality we’d be tempted to call aware. (Yes, I know this is a Dennettian sentiment also, but he’s just conveying a common outlook in the cognitive science communities.)

    A collection of syllables? Well, it’s difficult to communicate without using those 🙂 But for productive philosophical discussions, we have to be willing to make the effort to understand what the other person’s collections of syllables are trying to convey.

    My point was that eventually we get to a reflex, an action impulse. In primitive creatures such as c-elegans worms, this is close to a one to one relationship: stimulus A leads to action A, stimulus B to action B, etc. But with the rise of distance senses, the amount of information leads to situations where several reflexes might be triggered concurrently, and the organism has to be able to resolve which reflexes to allow and which to inhibit.

    In this view, the models are basically just cached stimuli, a mechanism to increase the scope of what the reflexes can react to. So instead of having to directly interact with a noxious or dangerous aspect of the environment, we have distance senses to give us early warnings of it. Perception (modelling), attention (allocation of resources), imagination (simulations), and introspection (self modelling) are ultimately all adaptations to increase the efficacy of our emotional reflexes.

  25. Hi Jochen,

    > Surely, to describe/interpret something, one must have a model of it?

    Yes, I would say so. It would be silly to try to interpret any system which is describing and interpreting something as not having a model of that thing. Any such sophisticated system is certainly best explained as being a model-using system.

    But it’s still a matter of interpretation. I would say that a computer program which interprets visual data as a 3D environment has a model of that environment. You might say it’s just pushing electrons and that it’s no different from a rock rolling down a hill in many-dimensional space. I would not go so far as to say that you are incorrect, but I would say that such an interpretation is not as useful as a model-using one.

  26. Jochen #21

    “The problem here is that you’ve swept all the difficult stuff under the rug in the first step—how does one get to having ‘models of the world’? How do, in other words, mental states come to be about the world at all?”

    Actually your question is a level.3 question. A fully general-purpose modelling system *includes* a model explaining how mental states come to be about the world.

    We start with the self-evident fact that there *is* a relation between mental states and the world, put that in the model as an initial assumption, and then refine the model recursively from there.

    The process of inference is part of the model of reality. So the answer is, check the model, it’s in there 😉

    Abduction:
    https://en.wikipedia.org/wiki/User:Zarzuelazen/Books/Reality_Theory:_Categorization%26Semantics

    Induction:
    https://en.wikipedia.org/wiki/User:Zarzuelazen/Books/Reality_Theory:_Probability%26Statistics

    Deduction
    https://en.wikipedia.org/wiki/User:Zarzuelazen/Books/Reality_Theory:_Symbolic_Logic

    3 types of reasoning only one of which (abduction) possibly has something to do with consciousness (induction and deduction can be unconscious as far as we know).

  27. A question about Karl Friston’s idea:

    Much inference is going on without our awareness or conscious thought. Non-conscious processes arise also from many complex systems. Why exactly aren’t they conscious?

    Does anyone have some thoughts?

  28. Hi ithio,
    Good question. I don’t recall Friston addressing it.

    My own thinking on this, at least currently, is that it depends on the type of modelling that’s taking place. We don’t seem to be conscious of passive modeling, but we do of the more active variety involved in imagination or concentration on a task. We seem to engage in that more active variety when imagining the past or future, or in novel situations, where mental habit (conditioned reflex) isn’t adequate.

    From a neuroanatomical standpoint, this makes sense, since both imagination (simulations) and introspection (self modeling) appear to be coordinated from regions in the prefrontal cortex (the front of the brain). Indeed, I suspect introspection is just part of the overall simulation system, but admittedly that’s speculative.

  29. Re 27 Ihtio.
    What we evolved from like other creatures, as well as trees and plants, lived without the need for consciousness. This established way continues today in the larger part of us which interacts with our smaller conscious part, which in turn decides on action, or not, in the environment in order to service the bodies requirements.
    Nature does not usually waste energy on what it can do without.

  30. Hi Peter,

    About Friston’s use of the word “inference”: he really does intend his readers to take it literally, not metaphorically. The nuance that needs to be supplied is that when he talks about inference, Friston is generally referring to an unconscious process. The aeon essay is a hand-wavy attempt to speculate about how inference in the usual conscious sense might eventually be incorporated into his account of how the brain processes sensory data. (I will get back to this.)

    The idea of unconscious inference goes back to Helmholtz (ca. 1870) who was able to measure the time it takes for a visual signal to reach the brain and demonstrate that there was a delay between the reception of the signal and the body’s reaction to it. Helmholtz concluded that the time lapse was necessary for the brain to make sense of the visual percept, something it does quite unconsciously as a rule. (Perhaps this is the origin of Schopenhaeur’s adding machine metaphor for the unconscious mind.)

    The idea of unconscious inference is uncontroversial in neuroscience. It led to the Bayesian Brain hypothesis (due to Hinton, I think) which is supposed to be implemented in the visual cortex by Predictive Processing.

    Unfortunately for Friston’s readers, standard accounts of Predictive Processing (PP) are not cast in Bayesian terms, although the translation is easily carried out:

    1. The generative model of visual data (the “prior beliefs”) is implemented in the backward network.
    2. Bayesian inference is carried out by the forward network. The job of the forward network is to calculate “posterior beliefs” about the causes of a given percept by minimizing “variational free energy”.
    3. The parameters of the forward and backward networks are jointly optimized, again by minimizing variational free energy.

    The PP framework can presumably be straightforwardly extended to accommodate passive perception of multimodal sensory data, not just visual perception. Friston’s nifty insight was that PP could be further extended to accommodate action as well as perception so that all sensorimotor activity could be accounted for in a coherent way.

    Of course a lot of work needs to be done to fill out this story, even in the case of an amoeba, let alone a brain. Not only does the state space need to be adequately specified (as you pointed out) but the “prior beliefs” about the state space dynamics need to be supplied as well.

    Suppose the amoeba has succeeded in carrying out Bayesian inference about the state of its world (body + environment). The Bayesian formalism requires that the amoeba now be able to enumerate all of the states of the world that can arise next and attach prior probabilities to each. Moreover, a really intelligent amoeba aiming to maintain or return to homeostasis would want to look several moves ahead rather than just one, so it would have to do some serious computing. My understanding is that this sort of lookahead is the “temporal thickness” that Friston is talking about in the aeon essay and this is where human access consciousness might (conceivably) enter the picture.

    BTW Friston’s suggestion that the Bayesian Brian is engaged in minimizing surprise has caused a ridiculous amount confusion which he could have avoided by carefully explaining the difference between entropy and cross-entropy.

  31. Thanks, Patrick. That tends to reinforce my earlier feeling that it’s rather difficult to get a clear idea of what he means from the Aeon piece alone.

Leave a Reply

Your email address will not be published. Required fields are marked *