The ‘Bayesian brain’ for beginners, by way of placebo

The Spectre of Rev. Thomas Bayes amongst the gyri and sulci

A recent Topical Review in Pain provides an interesting and gentle introduction to notions of Predictive Processing and the ‘Bayesian brain’ by way of medically unexplained symptoms and the placebo effect – two phenomena that have been notoriously difficult to account for in a purely biomedical model.

Click on image above for link to full paper

1. Introduction. The standard and ideal biomedical model of symptom perception treats the brain largely as a passive stimulus-driven organ. It embraces the notion that the brain absorbs sensory signals from the body and converts them, directly, into conscious experience. Accordingly, biomedicine operates under the assumption that symptoms are the direct consequences of physiological dysfunction and improvement is the direct consequence of the restoration of bodily function. Despite its success, the biomedical model has failed to provide an adequate account of 2 well demonstrated phenomena in medicine: (1) the experience of symptoms without pathophysiological disruption, and (2) the experience of relief after the administration of placebo treatments. This topical review advances the idea that “predictive processing,” a Bayesian approach to perception that is rapidly taking hold in neuroscience, significantly helps accommodating these 2 phenomena. It expands on recent high-quality empirical work on predictive processing and outlines, more broadly, how Bayesian models offer an altogether different picture of how the brain perceives symptoms and relief.

Read the whole Topical Review here.

If you want to dig deeper into Predictive Processing, two of the keys researchers have both written a series of superb, accessible blogs about their books on the topic – Andy Clark and Jakob Hohwy.

-Tim Cocks

Final 2018 NOI Course


2019 NOI Courses Australia and New Zealand

AUCKLAND 15 – 17 MARCH 2019 | EP+GMI

AUCKLAND  23 – 24 MARCH 2019 | MONIS





21 Responses to “The ‘Bayesian brain’ for beginners, by way of placebo”

  1. jqu33431quintner

    ” … how Bayesian models offer an altogether different picture of how the brain perceives symptoms and relief.” Can anyone make sense of this quote from the abstract of the Topical Review?

    • timcocks0noi

      Hi John
      I’ll make the same suggestion here as I did on Facebook – that one doesn’t throw out the Predictive Processing baby with the semantically inaccurate and awkward bathwater. I suspect that your issue is with the idea that the brain ‘perceives’ symptoms and ‘perceives’ relief, and i would agree that this sentence is essentially non-sensical. However the rest of the paper is an accessible introduction to Predictive Processing (originally Andy Clark’s phrase, distinguished from Predictive Coding in his seminal paper “Whatever Next…” ( and his book Surfing Uncertainty), arguably one of the more important current ideas (in its various guises including Karl Friston’s ‘Active Inference’ and Free Energy Minimisation formulations, Anil Seth’s notions on interoceptive inference and Jakob Hohwy’s prediction error minimisation approaches) in cognitive science.
      I’m not sure of the purpose of the seeming snarky attacks here and on Facebook? It’s easy to cherry pick a single sentence and pull it apart. If there are deeper issues with the predictive processing ideas why not engage with them? Surely this would be more illuminating than prickly one liners, no?
      My best

      • jqu33431quintner

        Tim, please understand that my comment related only to the final sentence in the Introduction. The link you initially provided did not enable me to access the full paper.

        As for Predictive Processing, I do not pretend to understand the construct, and therefore I am not in a position to disparage it.

        However, in my opinion the sentence does reflect badly upon the authors of the paper and upon those who reviewed it without calling it into question.

      • jqu33431quintner

        Tim, I have now had the opportunity to read the paper.

        The authors operate within the ‘you are your brain’ framework, which is quite acceptable in neuroscientific circles (but it is not a framework that I share with them).

        They attribute to the brain processes that previously would have been attributed to persons.

        For example, they assert the brain predicts, infers, forms hypotheses , but then go on to say: “To put it emphatically, we feel pain … because we predict that we are in pain.”

        I am puzzled as to why they have switched from “the brain” to “we”. Is there an homunculus dwelling within the Bayesian brain? The cartoon posted above is consistent with this interpretation.

        On Facebook I have raised some of the issues that I found to be confusing.

        The first author of the paper was kind enough to respond to my initial comment, as made above, and to my additional question on the definition of ‘symptom’.

        I am hopeful that he will continue to respond to my other comments and questions.

  2. jqu33431quintner

    Hi Tim.

    I have posted these extracts elsewhere, but perhaps you would also like to help clarify the meaning of each of them:

    “The predictive processing approach shows that [therapy] acts on the same inferential process whereby we experience symptoms relief”.

    “The interplay between descending predictions (sic) and ascending signals… is flexibly modulated by the ‘precision’ of hypotheses and sensory evidence”.

    “… all symptoms are product (sic) of an inferential process that is never strictly reducible to physiological dysfunction and is sometimes only loosely related or unrelated to it”.

    Thanking you in anticipation of your response.


    • timcocks0noi

      Hi John
      I can’t speak for the authors and wouldn’t attempt to clarify their words.

      Generally speaking, these sentences reflect ideas within the Predictive Processing framework, an exhaustive account of which would be outside the scope of a comments section! However, I did purposefully link to resources of a like kind (blogs) from two of the leading people in the field (Clark and Hohwy) for those interested in fleshing out the PP account.

      I think your quotes are misleading. For instance above you quote “The predictive processing approach shows that [therapy] acts on the same inferential process whereby we experience symptoms relief”. However the actual quote from the paper is “Crucially, the predictive processing approach shows that therapeutic ritual and active ingredients of the intervention, albeit through different routes, act on the same inferential process whereby we experience symptoms relief”. My reading of the actual statement is that the authors are emphasising the same underlying process (perception as inference) in explaining the effect of therapeutic ritual and active ingredients on percepts. If your concern is with the phrase’s semantic construction, i believe that the author (Giulio Oongaro) has explained this on Facebook.

      Secondly, rather than “The interplay between descending predictions (sic) and ascending signals… is flexibly modulated by the ‘precision’ of hypotheses and sensory evidence”, the authors state “Importantly, the interplay between descending predictions and ascending signals that lie at the heart of predictive processing is flexibly modulated by the “precision” (or “inverse variance,” in statistical terms) of hypotheses and sensory evidence”. This is a core idea within PP. The use of ‘hypothesis’ can be problematic as it is used metaphorically with the PP literature, but perhaps the broader context of the quote – the example of walking through a forest infested with snakes and perceiving a stick (sensory cue) as a snake (“hypothesis”), provides an explanation for the use of this terminology? Precision is a key concept in PP formulations but also has different meanings in more common use. The authors do make it clear that they are using the technical sense (inverse variance). Clark’s series of blogs do a good job of explaining the concept and its central role in PP.

      Finally, ‘… all symptoms are product (sic) of an inferential process that is never strictly reducible to physiological dysfunction and is sometimes only loosely related or unrelated to it’ misses the broader context again – “Importantly, the approach goes some way towards transcending the artificial but pervasive distinction between “explained” and “unexplained” symptoms in biomedicine (or
      between “real” and “imaginary” illnesses). One important upshot of the theory is that all symptoms are product of an inferential process that is never strictly reducible to physiological dysfunction and is sometimes only loosely related or
      unrelated to it’. I can’t see the confusion here. Far from denying the phenomenology of an individual’s symptoms (as you have suggested on Facebook) these statements object to medicine’s “artificial” categorisation of symptoms unrelated to identifiable pathology as “imaginary”. And further, that even in the presence of physiological dysfunction, one can not say that the lived experience, the phenomenology, just is these dysfunctions.

      There are some wording difficulties in the paper – Giulio has said as much himself in a reply to you on Facebook, but the broad thesis is still one that deserves consideration, i think.

      My best

      • jqu33431quintner

        Thanks Tim. I agree that the broad thesis is deserving of consideration (and discussion) in relation to the experience we call pain. However, I think that some of the philosophical and psychological implications are important and transcend the wording difficulties in the paper. They will, I hope, be raised in a Letter to the Editor.

        • timcocks0noi

          Indeed! Far from consensus, the realm of PP is full of rigorous debate and deep disagreement. Hohwy and Clark are seen to be at opposing ends of a spectrum regarding how ’embodied’ PP may or may not be (but in fact are probably closer to each other – personal communication), another reason for linking to both.

          I would welcome, and i would hope that the authors would welcome, a letter to the editor to further enable this discussion.
          My best

  3. Adam Bjerre


    The (unintentional) move you identify is also called equivocation: the use of a term that is open to two or more interpretations, hence a potential source of misleadingness. Poets and con artists may seek equivocation, but in law and in science the choice of words are aimed at more narrow interpretations to constrain the concepts and uphold scientific rigour.

    Plenty of terms in cognitive science, AI and biology are used in this way, invoking a kind of cryptic dualism, the tendency to treat phenomena interchangeably as purposeful and nonpurposeful. Terms like mechanism, information, coding, template, representation, aiming, control, cause – are often used two different ways, either as though they’re functionless or as though they’re functional. One could argue that anticipation, prediction, and inference are used the same way – unintentionally hiding or explaining away the homunculi that they depend on.

    The reason this move is so ubiquitus in these fields is probably the problematic relationship we have with end-directedness, telos. To get away from black box explanations and circular reasoning that teleology traditionally invokes, we are continuing in the Cartesian tradition with bodies as machines, brains as computers and animals as “beast machines”.

    Jessica Riskin has highligthed the centuries long argument over the banning of agency in the life sciences in The Restless Clock.

    • jqu33431quintner

      Adam, thanks for your reasoned response. But can you tell me why some of the academic physiotherapists are so excited by this paper?

      • timcocks0noi

        Perhaps you are reading too much into this John? I’m not aware of any particular excitement amongst ‘academic physiotherapists’ (whoever they may be!?) regarding this paper.
        I posted it because it relates to Predictive Processing (a topic that interests me and that there is considerable excitement about – see Andy Clark’s ‘Whatever next’ paper and the many responses from a range of luminaries in the philosophy of mind and cog sci world as just one example) and pain, and was nominally ‘open access’.
        I do hope that physiotherapists (and other health professionals) get excited by topics such as these (phil of mind, phenomenology, cog sci) in order to both progress our understanding of human experience, and realise the severe limitations of our current thinking. Notwithstanding that there will be bias (my particular interests already declared) disagreements, discussion and debate along the way.
        My best

        • jqu33431quintner

          Tim, I have re-read the paper and could not find any mention of “imaginability” – a mental phenomenon that I thought would have been a key component of any adaption of predictive processing theory to the clinical context. How else but by imaginative construction of meanings, interpretations, and predictions, can clinicians and their patients expect to negotiate effectively in the course of their therapeutic engagement? Have I missed the point?

    • timcocks0noi

      Hey Adam
      It’s been a while! Thanks for stopping by.
      Perhaps there is another option too – where a word is used with great care and precision (such as the word ‘precision’) but has a very definite and different meaning within the context of a certain domain.

      After Dennett, perhaps there can be predictions without a predictor? Friston might add that the very morphology of an organism is, in itself, a ‘prediction’ of the kinds of environments that the organism is likely to find itself in. Of course Friston not only acknowledges, but embraces, the glorious circular reasoning and causality inherit in these arguments!

      My best

      • Adam Bjerre

        Hi Tim,

        Yeah, it’s been a while. Digesting this stuff on the side of full time work and family is a wonderful hobby, but takes time (and is a bit of a struggle sometimes) for a non-academic.

        I think you’re right that the words and terms mean or represent different things depending on the domain. The words we use often carry a lot of baggage, so we have to be very careful with our wording and choice of terms.
        I often see a conflation of the word information used in the technical sense (difference, bits-and-bytes, Shannon-information) with information used in the normal sense (about something for someone, aboutness, normative) without the assumption being addressed (see esp. Clarks BBS-paper).

        My guess is that Fristons notion could also be characterized as functional fittedness. The interesting question is what kind of logic accounts for functional fittedness. Is self-organization, i.e. regularization processes, enough?

        Regarding Dennett and predictions without a predictor, I recall a podcast he did with Sam Harris, where they talk as though it has been established that organisms/agents/subjects/creatures/living beings/selves, or whatever we want to call them, are just cause-and-effect machines, or fancier computers, and the only remaining question is how to break it to us.
 Cause-and-effect mechanism and prediction/anticipation/inference/interpretation/representation/aboutness is not the same though the latter depend on the former, as I understand it (so no place for any kind of supernatural forces, added vital powers, or whatever). And it has nothing to do with whether it is a wonderful or dreadful issue, as Dennett often insinuates. The critical arguments I have read in developmental biology centers around the crucial point that it is logically incoherent to claim that we are coding machines, that biomolecules are little “stupid” machines, or that the brain is a coding or processing engine.
        There is a fundamental difference between engineering logic and biological logic. They are opposites. Engineering logic start out with parts that gets assembled to do some function, biological logic start out with functional wholes that differentiate, like sculpting. We need to establish what a self is, a system that can self-regenerate, before we can establish what predictions are about for whom. The reverse-engineering fallacy is the assumption that science can be done in the same order as engineering – prescription though explanation toward description and not the other way around as is normally or ideally the case in science.
        We are neither ghosts nor machines, as Jeremy Sherman aptly puts it in his book from last year.

        I’m glad to be corrected if my interpretation is wrong-headed, which very well could be the case for an amateur. :-)

        “Truth spring from argument amongst friends.”
        – Hume

        Thanks for your careful and helpful work on this site, Tim.


  4. Adam Bjerre


    No, I can’t tell you why. That would be wild speculations. Maybe trust in the authors or the promotors of the paper? Maybe because the ideas are gaining a lot of support in the field of neuroscience? Maybe because the paper continues the tradition of finding correlations between cause-and-effect relationships and means-to-ends behaviour, looking for neuronal mechanisms that produces predictive processing?

    As I understand it, it is almost unquestionable in both biology and cognitive science that organisms predict or anticipate. Hence it could be argued convincingly that perception is constrained hallucination regarding us humans. It is the basis of adaptation or learning.

    But we still lack a theory of the physical nature of predictions/anticipations/inferences/interpretations. Not that it will never be possible. I think actually we may be closer to an explanation via the insights from complexity theory, dynamical systems theory, information philosophy and evo-devo.

  5. jqu33431quintner

    Tim, at Giulio’s request, I would like to transfer my comment on Facebook to this blog. He has undertaken to respond.

    Giulio, perhaps you would be kind enough to correct me if I have misinterpreted the following argument in your paper: ‘Given that the same inferential process is implicated in both cases, the theory also explains why the so called “real” and “imaginary” symptoms seem to be phenomenologically indistinguishable from the patient’s point of view.’? As it reads to me, the inference being made here is that until such time as the physician can explain a patient’s pain in terms of discernible ongoing nociception, it is classed as “imaginary”. The reason I raise this point is that pain sufferers not infrequently complain they are told by their health care professional that their pain is “imaginary” or “unreal”.

    • Giulio Ongaro

      Hello. Sorry for the delay in replying (just had a very hectic week). I’m happy that the post has sparked some discussion.

      I would subscribe to a lot of what Tim has said in response to John’s misgivings, especially on the three quotes you took for the text above. I would have said pretty much the same thing.

      As for your latest doubt about the ‘imaginary’: I am not sure I understand the question very well. You seem to imply, from your reading of the paper, that there will be a time when the physician can explain a patient’s pain in terms of discernible ongoing nociception. But the whole point of predictive processing is to say that all types of experience stem from internally generated predictions about the state of the world and body. In a sense, all pain is ‘imaginary’. What differs between what is usually classed as ‘imaginary pain’ and ‘real pain’ is the degree to which the experience is coupled to nociceptive inputs. Clearly, acute pain is primarily determined by nociceptive input, but even here, even in the most basic instances of acute pain like touching a hot stove, the experience originates from internally generated predictions about the state of the body. Very simplistically put: until I touch the hot stove my phenomenological state is determined by ongoing top-down predictions of being painfree. As a touch the hot stove, sensory inputs deviate from the predicted state. To minimize prediction error, the brain generates a prediction of being in pain that match the input. Now, as Tim said, the concept of ‘precision’ in central in all this. The brain is in the constant business of weighting the precision, or reliability, of bottom-up inputs and top-down predictions depending on context. I.e. in order to infer the state of body and world the brain estimates whether it should rely more on sensory inputs or on internal predictions, which have been informed by previous experience (both ontogenetic and phylogenetic). In the case of chronic pain the balance shifts towards predictions, which, highly weighted, ‘quash’, so to speak, the sensory evidence. In cases of chronic pain, even small interoceptive changes that in normal individuals would hardly register prompt the brain to infer pain as the cause of these changes.

      I admit that all this is very unrefined, that there are a lot of philosophical issues that cry out to be addressed. This is due to the space limit. One of the issues you have picked up is that a lot of people who write about predictive processing, including myself in that short paper, engage in the ‘you are your brain’ kind of talk, and conflate ‘brain’ with ‘we’. This move is problematic, but I think we can get around it with some philosophical work (some philosophers even see the theory as ultimately externalistic!).
      It’s important to understand, though, that the conceptual baggage employed by predictive processing theorists – prediction-signal’ ‘prior’ ‘prediction error’ ‘likelihood’  ‘input’ etc – isn’t representationalist. Specifically, these terms refer to to probabilistic distributions in neuronal populations. It’s not ‘mentalese’ language. Quite a few philosophers have picked up on the non-representationalist nature of predictive processing, and argued for the compatibility of phenomenological/enactivist and Bayesian theories. You might want to look for instance at this paper by Adrian Downey:

      We didn’t have space to address any of the above in that tiny paper. Our main objective was to demystify ‘medically unexplained symptoms’ and ‘placebo effects’; to show that these are just particular manifestations of the ordinary way the brain works. I think the thrust of predictive processing does not lie in offering another account of puzzling phenomena like MUS and ‘placebo effects’, but in radically upending the conventional understanding of the phenomena that we thought were self-evident or easy to account for, like acute pain and treatment effect, and show that there is a continuum between these and MUS and ‘placebo effects’, as explained, albeit tentatively, in the paper. We think this has important implications for medical practice. At the very least, it should prompt medical practitioners to rethink what is often dismissed as ‘imaginary’ pain, to appreciate its reality if you like, so that patients are not left with the impression that they are making up their pain (which you mentioned is a widespread problem). Of course, how exactly one should implement a better medical practice based on this understanding is a different matter.

      I liked the points raised by Adam in the other post. I am myself sympathetic to the critique of the machine metaphor in cognitive science. I haven’t read Jeremy Sherman’s new book but I was very much taken by Terrence Deacon’s “Incomplete Nature” (on which Sherman builds) as well as by other ‘enactivist’ anti-computational literature (e.g. Evan Thompson’s “Mind in Life”). Yet I see a large degree of compatibility between these and predictive processing. Predictive processing is still in a state of infancy and will be highly reformulated in the years to come. But I like to think that the central idea that perception stems from internally generated predictions about the state of the world constantly reined in by sensory evidence – that, as Seth put it, perception is a sort of ‘controlled hallucination’ – is bound to stay.

      Why is this exciting? Asks John. Well, if you read earlier versions of how the mind works (e.g. Steven Pinker’s “How the mind works”) you realize that people used to have (and many still have) a very different idea of how the mind perceives the world and acts upon it. PP offers a radical alternative that in my view does full justice to the role of context and culture in perception, and is increasingly backed up by a wealth of experimental evidence.

      I can’t claim I’ve been entirely clear and exhaustive in this post. Partly, though, this is because predictive processing itself is a very complex theory, which is hard to distill in a few lines. In that wee review, Ted Kaptchuk and I hoped to have given at least a hint of what’s at stake in it in relation to pain.
      For a comprehensive account that I am sure addresses your concerns more satisfactorily I’d recommend Andy Clark’s “Surfing Uncertainty” (I was one of his Msc students in 2013 when the ideas in this book were still brewing). A number of philosophical issues surrounding PP are tackled in this collection:

      I hope this helps!

      • jqu33431quintner

        Guilio, thank you for responding to my concerns about your paper, some of which have not been allayed.

        By way of a final comment, in your paper you have not mentioned the pivotal role of natural language, which according to philosopher Horst Ruthrof [2014] can be defined as a “social set of instructions for imagining, and acting in, a world.”

        As he points out, “meaning could be said to occur when language users are able to imagine relevant mental scenarios in response to linguistic expressions.”

        Accordingly, I can readily imagine that you are experiencing pain, but I cannot imagine that I am experiencing pain only because of my brain’s prediction that I am in pain “based on an integration of sensory inputs, prior experience and contextual cues”. The predictive processing model of pain makes no sense to me, but I accept that others will disagree.

        Reference: Ruthrof H. Language and Imaginability. Newcastle upon Tyne: Cambridge Scholars Publishing, 2014: 236.

        • Giulio Ongaro

          Well, a major point in the paper was that predictions are also formed on the basis of contextual cues, which encompass linguistic expressions.
          In fact, I do mention the role of ‘verbal suggestions’, ‘verbal information’ and ‘verbal interaction’.

          • jqu33431quintner

            Giulio, what are the contextual cues that I might be using when I predict that my experience will accord with that of being in pain?

            Do I need an unbiased observer to tell me whether I should or should not be having such an experience?

            In your paper you do not tackle this important issue, apart from stating that to the experiencer they are phenomenologically indistinguishable. Why then do you make the distinction?

            Finally, should the observer decide that my pain is not “real” but “imaginary”, does that imply I did not have such an experience and that it was an illusion?

            I labour these points because Tim has recommended your paper on the “Bayesian Brain” as being suitable for beginners. As a beginner, I find it most perplexing, to say the least.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

Basic HTML is allowed. Your email address will not be published.

Subscribe to this comment feed via RSS

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: