I'm full of the most awful cold at the moment so that might be to blame for my mental fog, but, if you've time, I think I might need you to lay out (C) in a little more detail for me. — Isaac

Not( Paper => (C) )

out of garbled logic

is

The paper doesn't impact the question:
Are there hidden states which we don't access through the sensorium (through innate bodily capacities) which we access through other (ideational, tool based) means?

We don't have a sense for alpha particles. But we have geiger counters and know about radiation.

I pointed this out because:

(Perception model) to (Inherent constraints on the thinkable/knowable)

is all too common.

wouldn't radiation still have a function with regards to its causal relationship with environmental states? — Isaac

I think the paper is silent on whether the hidden state vector $\gamma$ is a complete specification of a body's causal relationships with its environment. Stuff like radiation existing makes me doubt that it would be; if Friston's model is right, it's always been; we've always been radiation sensitive, but for most of our history we've not known about it, understood it, intuited it, or been able to detect it.

Ahh. So the idea is that we might be able to construct linear-functionally related models (with or without noise?) of hidden states which bypass the function which relates those states to perception because we're no longer perceiving them directly, but through machines, and machines don't infer causes before reporting then, they just report them directly.

This is good because it gets to the heart of model-dependant realism as Hawking expresses it (he's talking almost exclusively about models in physics - like radiation, of course). It was the relationship between Hawking's idea and Friston's which interested me.

I agree with you that Friston's paper doesn't impact on that debate any further than to say that, if you buy model-dependant realism in physics, here's a good way to look at it in organisms too.

Apart from...

What are we to say about the observation of the reading from the Gieger counter? Obviously the actual sensory input is in terms of black numbers on a white background, but the reason that a certain shape becomes a 'dog' is because we have a model of 'dog' which somewhat matches a good possible cause of sensing that shape. Can we not say the same about the Geiger counter? What's the limit to the extent to which our inference of cause extends back outside the Markov blanket? Obviously we've no free-energy incentive to model anything more than one node outside the Markov blanket, but maybe, like with so many evolutionary traits, we've gone into overclocked mode with our inferences?

If any of this is helpful, then the next step is whether these non-perceptually constrained inferences have any necessary accompanying increase in the accuracy with which they match hidden states (noise reduction), and, more importantly, whether they match them with any exclusivity to other possible matches. My feeling is "yes" to the first, but "no" to the second.

Obviously we've no free-energy incentive to model anything more than one node outside the Markov blanket, but maybe, like with so many evolutionary traits, we've gone into overclocked mode with our inferences? — Isaac

It might also be that we can create states that are sensitive to hidden states outside of our Markov blanket! So there might not just be "one blanket", and it becomes activity and history dependent. We might be able to engage in actions, like reasoning and concept creation, that let us "peer beyond the veil" of our immediate bodily constraints, and internalise those concepts and practices to learn how to sense and act inferentially in new ways.

If I looked at a graph I'd be able to interpret it in some way, if I showed the graph to someone 10,000 years ago; they'd have no idea.

So there might not just be "one blanket", and it becomes activity and history dependent. — fdrake

Just reminded me of a paper by Peggy Series (I think) about nested Markov blankets and how they can be used to create wider network models - something like the 'extended brain' idea. It think it seems a very promising proposition.

nested Markov blankets and how they can be used to create wider network models — Isaac

I guess fundamentally the point of relevance is that humans are not just recipients of hidden states, we are creators of them - active perception both interfaces with present and creates new states.

I guess fundamentally the point of relevance is that humans are not just recipients of hidden states, we are creators of them. — fdrake

Probably only tangentially related (but then isn't everything we've been discussing - what was the actual thread topic again?). I did some work many years back on the unpredictability of social networks, how they create emergent states (strange attractors) for certain variables in a random way. Like a very long game of Chinese whispers with the players standing in a circle, after several thousand iterations the word is so far removed from it's original, it's almost unrecognisable, but, it's still a word, it's been constrained by the morphological field of language, not by the actual individuals.

Anyway, I was looking at the many variables of a social group which could be said to have emerged this way, rather than to be an actual intentional feature of any individual(or group of individuals).

I don't know if I'm just over-thinking it, but it seems possible to see these variables as 'hidden states' because none of us directly know what their causes are (we didn't deliberately make them), but they are nonetheless entirely created by the social group trying to infer what they are.

what was the actual thread topic again?). — Isaac

Boring. It was boring.

, but, it's still a word, it's been constrained by the morphological field of language, not by the actual individuals. — Isaac

I've done some free time coding looking at equilibrium distributions of random variables on dynamical random graphs; in these contexts the initial state ("what was said at the start of the game of whispers") don't matter, what matters for the equilibrium state is the action (the whisper mechanism) that iterates the random variables (the words) associated with each node (the people) and the architecture of the connection of the graphs (who speaks to who).

I dunno how this would relate to strings being passed around graphs, as I've just used numbers. I imagine that it's possible to think of the space of strings as a metric space under a string distance metric (like letter mismatch count in a simple case), and the "whisper action" takes an input string (say "poop") and outputs a string "say "boob") with probabilities as a decreasing function of the distance metric.

I have an intuition that strings are "sticky" in the whisper action in a way numbers aren't typically: if an input whisper was "boob" and then the output whisper was "beeb", the back transition is equally as improbable as the forward transition, but still possible, so after an improbable transition (mishearing oo to ee), the whispers are likely to stay closer to "ee" sounds in the middle of the word for a while.

Edit: Could probably say the same thing about "large jumps" in a continuous quantity when the action is sufficiently a "local jump". Dunno how good this intuition really is. I'm imagining that it comes from the state space of strings being discrete? But how sticky it is is ultimately just a question of how "heavy tailed" and "has most of its mass close to 0 distance changes" the transition kernel would be.

I don't know if I'm just over-thinking it, but it seems possible to see these variables as 'hidden states' because none of us directly know what their causes are (we didn't deliberately make them), but they are nonetheless entirely created by the social group trying to infer what they are. — Isaac

We can probably do something like that; say there are two different whisper actions, ways that people pass words around the social network with some error; we could say that two different whisper actions are equivalent on a set (of string space) A with degree p when and only when the union of the attracting sets has A in proportion p.

Like if we imagine two dynamical systems tracing out the letter "o" in space close to each other with some error, so that the overall evolution looks a lot like the number "8" (one "o" above the other"), the bridge in the middle of the 8 would be an A and the proportion p would be the total area of the part of the union of the attracting sets around the bridge between the upper and lower components of "8".

A has a random variable associated with it, I( A ), the indicator function on the union of the phase spaces on the dynamical systems that is 1 when the current point of both is within A and 0 otherwise. This is a "hidden state" being "partially explored" by the "measurement process" of the tracing states, but so would be any random variable associated with the union of the phase spaces.

Interpreting this causally would be difficult. The overall causal structure is embedded in the dynamical system actions; transition probabilities on states; rather than random variables defined as summary characteristics on the states.

the initial state ("what was said at the start of the game of whispers") don't matter, what matters for the equilibrium state is the action (the whisper mechanism) that iterates the random variables (the words) associated with each node (the people) and the architecture of the connection of the graphs (who speaks to who). — fdrake

Yes, this is exactly what we found, initial states are largely irrelevant. What I really want to do some day (but semi-retirement makes it an increasingly unlikely prospect) is to actually set up such an experiment. I had all these ideas about how we could do it with words, actions, I even had a way of trying it with opinions (comparative beauty judgement was my tool here).

The idea would be to see exactly what you were intimating about "beeb" and "boob", whether there were attractors which, once a sufficiently large error triggered a move to, were reluctant to leave. My hypothesis was that there would be such attractors and that they would tell us something interesting about the sort of error correcting mechanisms we use. I didn't know anything about Friston at the time, so my idea of error correction was far too linear, and should have been more probabilistic, but my instinct was that we'd have a method of 'correcting' what we thought were errors in communication which were themselves biased toward some presumption. With language that would be phonemes "I expect that was an 'ee', most words are", with opinions...well, that was going to be the interesting bit. Your statistical parsing is very helpful in pinning down how these connection might look in actual results.

The idea would be to see exactly what you were intimating about "beeb" and "boob", whether there were attractors which, once a sufficiently large error triggered a move to, were reluctant to leave. — Isaac

I'd be interested in hashing out an experimental design with you at some point if you'd be interested.

This is the crux of it. Without input the brain lacks a means to locate itself within an environment. Sensory deprivation produces ‘hallucinations’ because we’re essentially rummaging through our mental tool box and randomly applying tools that may help us navigate.

@fdrake At least I can see where this discussion turned from philosophical approaches to neuroscience. It is a little confusing when two related, yet altogether different discussions are going on parallel in one thread.

There's a distinction which I either keep failing to explain properly, or people don't generally seem to think useful, but it's crucially important to model-dependant realism, that is between reality having structures and reality being composed of the structures we divide it into.

I've used this example before, so apologies for the repetition if you've been following the whole thread, but it's like the constellation Orion. It definitely is in the shape (vaguely) of a hunter with his bow, belt and dagger, it's not that such structure isn't there, but it's also on the shape of just about anything else you could draw between those points, maybe not an infinite number of things (I'm not myself sure on this point), but certainly more than the one structure we impose on it out of that range of possibilities.


So to your point about reality having structural regularities which are 'real', yes, I think such regularities are not only only real, but necessarily so. If reality were homogeneous there would be no random direction to entropic forces and so no probability gradient against which the free-energy reduction would work. What I don't see is any reason why those structures must exist uniquely defined. So when you say "wavelengths picked up by the retina are coming from reality" I don't think there's any reasonable way we could disagree, but 'wavelengths' are themselves a concept, they're just one way of dividing energy among others. We can't even determine if wavelengths are a wave in a field or a particle, not that we've 'seen' either because both are just models interpreting numbers on a computer (which are the only thing we actually have 'seen'). — Isaac

I think the bold is relevant here. The constraints on what possible structures you can define matter as they point out invariancies that are relevant for structure. I'm thinking now back to my rabbit-duck example from earlier. You could imagine many ways to conceive of it other than either rabbit or duck. Maybe take the rabbit ears as handles for a new tool, with the neck of the rabit as the actionable end of the tool. etc. But you certainly could never see that figure as a circle or sphere. Those geometric relationships are conserved over all possible ways of perceiving the object.

In the case of a physical system it's the same idea, you have some constraints on how things operate - you, for example, will always see a diffraction pattern if you shoot an isolated electron through one slit. Or, taking the Orion star example, there are a finite number of ways to connect the stars and create a structure out of the set of stars constituting Orion. The ways of connecting the stars and the orientation of the edges between those stars, are constrained by the relative spatial locations of each star. Going back to the wavelength example, there is a power distribution to the frequencies of light which constrain the possible set of color experiences any observer could have. Even if you had an IR sensing organism, if the power in the IR bandwidth is low, that organism would not see that object and so on.

My point is there is structure that I think is more than merely random heterogeneity of properties over a space. Rather you have patterned heterogeneity, and those patterns are dictated, ultimately, by the structure of the space and the relations between the underlying physical properties.

Another metaphor might be to think of reality as a multi-dimensional contour map, it definitely has hills and valleys (ie it definitely exists and had variable structures), but which dimension should take precedent in determining what features are 'hills' is an arbitrary decision, or in our case, probably a pragmatic one limited by the biological hardware we've managed to evolve.

I agree, but I think at a certain level or, perhaps in general, there is some constraining set of rules that define reality primitives. And I say this because enduring structures are possible in the first place.

Really interesting point about reaching indeterminism in our models and what that means for how fundamental they are. I'm tempted to agree with you that indeterminacy cannot be further reduced, and so if we had it right this would not be one-pattern-among-many but would truly be the entity out of which patterns are made (like finding the actual stars in my Orion example). I'm wary to commit to it though because we'd have to remember that all this is within one huge Ramsey sentence about quantum physics, the first 'If' of which may well be wildly off mark.

Well I'd hope it isn't, but given how many holes there are in the standard model, maybe that's something to really worry about. It would be so depressing to think all of our 'advancement' in empiricism led us to left field when the ball was going right. I mean that would just be shattering for me lol. It's what drives my science interest.

What's fascinating about indeterminacy at the heart of the whole thing is that it might make our estimates of noise truly Gaussian (rather than just the assumption of Gaussian in our models) by the , at a fundamental scale, which is a point I think fdrake made about central limit theory.

Well it's interesting because probability distributions differ, thinking of electron orbital shapes, of interferometer experiments where there’s a non 50/50 likelihood for the particle to land at either detector and so on. What sets those is a complete mystery to me.

My point is there is structure that I think is more than merely random heterogeneity of properties over a space. Rather you have patterned heterogeneity, and those patterns are dictated, ultimately, by the structure of the space and the relations between the underlying physical properties. — aporiap

I don't doubt there are limits (to doubt that would lead to idealism), but I doubt we could ever describe those limits, we can only refer to them tangentially by pointing to ineffective models and speculating that transgressing one of those limits may be the cause of its failure.

Essentially, it's the problem of pessimistic meta-induction. We cannot reasonably induce that our theories model reality with some one-to-one relationship because absolutely all the evidence we have from previous models is that our models do not do that. If we were to speculate that our current models reflect reality in some unique way (by which I mean not merely one of a number of equally viable options), then we'd be faced with an explanatory gap as to why these particular theories have such a relationship when clearly every single past (rejected) theory did not. The evidence is overwhelmingly in favour of the conclusion that our current crop of models will go the same way in time.

It would be so depressing to think all of our 'advancement' in empiricism led us to left field when the ball was going right. I mean that would just be shattering for me lol. It's what drives my science interest. — aporiap

Indeed, but philosophically more interesting maybe... That might be some compensation.

Well it's interesting because probability distributions differ, thinking of electron orbital shapes, of interferometer experiments where there’s a non 50/50 likelihood for the particle to land at either detector and so on. What sets those is a complete mystery to me. — aporiap

That's interesting. Do we have, as part of the model, the factors affecting the distribution, or is that part of the mystery?

I don't doubt there are limits (to doubt that would lead to idealism), but I doubt we could ever describe those limits, we can only refer to them tangentially by pointing to ineffective models and speculating that transgressing one of those limits may be the cause of its failure. — Isaac

Why do you think we can’t describe the limits— it seems to me they can be describable, as I gave examples of in the earlier post with duck rabbit and Orion. The geometric relationships are the limit.. no matter the theory of ‘what’ those geometric relationships represent [duck, or rabit], the geometry is invariant. Other kinds of things which don't have those geometric figure boundaries are not representable. What is your take on structural realism?

Essentially, it's the problem of pessimistic meta-induction. We cannot reasonably induce that our theories model reality with some one-to-one relationship because absolutely all the evidence we have from previous models is that our models do not do that. If we were to speculate that our current models reflect reality in some unique way (by which I mean not merely one of a number of equally viable options), then we'd be faced with an explanatory gap as to why these particular theories have such a relationship when clearly every single past (rejected) theory did not. The evidence is overwhelmingly in favour of the conclusion that our current crop of models will go the same way in time. — Isaac

Pessimistic meta induction implies no carry over between theories... which to me seems clearly wrong. When electrons were first thought to be particle like, and then recognized to have wave properties when isolated, the new wave-particle theory didn't completely do away with the previous laws describing their motion and properties, it subsumed them. Newtonian laws which describe motion, while originally thought to be universally applicable to objects of all sizes moving at all speeds, is not done away with but subsumed by Einstein's relativity theory, and considered consistent with it given specified conditions. And so on.

Indeed, but philosophically more interesting maybe... That might be some compensation. — Isaac

I can't think of another answer generating method, and answers is what makes me satisfied; the irony of my username I guess. Philosophy is a sieve for ideas and generator of possible explanations but not of plausibly definitive answers.

That's interesting. Do we have, as part of the model, the factors affecting the distribution, or is that part of the mystery? — Isaac

I wish I knew enough to say.

Why do you think we can’t describe the limits— it seems to me they can be describable, as I gave examples of in the earlier post with duck rabbit and Orion. — aporiap

True, but with the Orion and duck/rabbit examples we are able to talk in the meta-language about the matter from which they're constructed. That's what enables us to 'know' the boundaries. What would happen if, for example, we became able to see in infra-red and ultra-violet. We see those wavelengths just as we do normal colours. We then look at the duck/rabbit and see a pig also, but one drawn cunningly in only ultra-violet and infra-red. Now where's our certainty that only a duck or a rabbit are possible?

What is your take on structural realism? — aporiap

I can get on board with Ramsey style epistemic structural realism, but not the traditional version. There are a number of objections to structural realism of the more traditional kind and I admit that some of them are over my head, I'm no mathematician, but the one I think I do get is that we have not been able to demonstrate that - even if the mathematical relations of a previous theory acted as bounds to all subsequent ones - the mathematical language we're using is actually responsible for (rather than incidental to) the theory's success. This is the point Stathis Psillos makes, I think.

As a means of focussing new theories, I think it's a great way of looking at realism. As an actual answer to redeeming scientific realism unscathed, I'm not so sure.

When electrons were first thought to be particle like, and then recognized to have wave properties when isolated, the new wave-particle theory didn't completely do away with the previous laws describing their motion and properties, it subsumed them. Newtonian laws which describe motion, while originally thought to be universally applicable to objects of all sizes moving at all speeds, is not done away with but subsumed by Einstein's relativity theory, and considered consistent with it given specified conditions. And so on. — aporiap

And so on indeed, but only for theories expressed in mathematical terms already. Note you've not included any theories of biology, psychology, even chemistry there. Mathematical structure may be preserved in theories which are expressed in that form, but there's no evidence it is in theories not expressed that way and so it still remains that structural consistency might be an artefact of the means by which we describe, not that which we describe.

Philosophy is a sieve for ideas and generator of possible explanations but not of plausibly definitive answers. — aporiap

Yes, I'd agree with you there.

I wish I knew enough to say. — aporiap

A feeling I have most often.

True, but with the Orion and duck/rabbit examples we are able to talk in the meta-language about the matter from which they're constructed. That's what enables us to 'know' the boundaries. What would happen if, for example, we became able to see in infra-red and ultra-violet. We see those wavelengths just as we do normal colours. We then look at the duck/rabbit and see a pig also, but one drawn cunningly in only ultra-violet and infra-red. Now where's our certainty that only a duck or a rabbit are possible? — Isaac

I don't see a problem with conditional rules [in this case of duck/rabbit: if restricted to visible light spectrum, then the limits are the geometric boundaries of the figure]. I would still consider them objective in the sense of mind-independent, but you could never guarantee their stability over time. E.g. the cell theory thesis that cells are the smallest unit of life is clearly only applicable to terrestrial life within the timespan of life's existence on earth. But it is still objectively true in this domain.

Physics, would be the one field I think there would be possibly unconditional rules. When, for example, you identify frequency and wavelength as properties of light, those must be describing some features of reality that are invariant to theory change. Even if we discover new properties, or incorporate wavelength and frequency in some other conceptual scheme of light in the future, those features would still be components of the underpinning reality, albeit understood in terms of additional relationships to other properties.

I can get on board with Ramsey style epistemic structural realism, but not the traditional version. There are a number of objections to structural realism of the more traditional kind and I admit that some of them are over my head, I'm no mathematician, but the one I think I do get is that we have not been able to demonstrate that - even if the mathematical relations of a previous theory acted as bounds to all subsequent ones - the mathematical language we're using is actually responsible for (rather than incidental to) the theory's success. This is the point Stathis Psillos makes, I think.

As a means of focussing new theories, I think it's a great way of looking at realism. As an actual answer to redeeming scientific realism unscathed, I'm not so sure. — Isaac

This is what I am essentially defending is a ramsey style ESR. I don't think there would ever be a way of verifying whether structure is all there is, as in OSR.. so I think it is a bit too extreme. And non-ramsean ESR may reach too far in trying to justify theoretical entities without recourse to concrete referents... But I certainly think some version of structuralism works here. I think the no miracles argument against antirealism coupled with the predictive power of empirical theories are the strongest arguments against pure antirealism

And so on indeed, but only for theories expressed in mathematical terms already. Note you've not included any theories of biology, psychology, even chemistry there. Mathematical structure may be preserved in theories which are expressed in that form, but there's no evidence it is in theories not expressed that way and so it still remains that structural consistency might be an artefact of the means by which we describe, not that which we describe. — Isaac

I'm curious what you mean by structural consistency being an artifact of the means by which we describe, and not what we describe. Wouldn't you say explanatory equations are derived from the empirical process? To know, for example that F = mass x acceleration, you must observe that the magnitude and direction of force is the multiple of mass and acceleration. To know the fact that the rate of a reaction is proportional to the concentration of the reactants, you need to observe it empirically.. and so on. I think of mathematics in science as no more than a language for precise expression for well-defined observables.