What do you think?
Rate this book
Surfing Uncertainty: Prediction, Action, and the Embodied Mind
How is it that thoroughly physical material beings such as ourselves can think, dream, feel, create and understand ideas, theories and concepts? How does mere matter give rise to all these non-material mental states, including consciousness itself? An answer to this central question of our existence is emerging at the busy intersection of neuroscience, psychology, artificial intelligence, and robotics.
In this groundbreaking work, philosopher and cognitive scientist Andy Clark explores exciting new theories from these fields that reveal minds like ours to be prediction machines - devices that have evolved to anticipate the incoming streams of sensory stimulation before they arrive. These predictions then initiate actions that structure our worlds and alter the very things we need to engage and predict. Clark takes us on a journey in discovering the circular causal flows and the self-structuring of the environment that define "the predictive brain." What emerges is a bold, new, cutting-edge vision that reveals the brain as our driving force in the daily surf through the waves of sensory stimulation.
In this groundbreaking work, philosopher and cognitive scientist Andy Clark explores exciting new theories from these fields that reveal minds like ours to be prediction machines - devices that have evolved to anticipate the incoming streams of sensory stimulation before they arrive. These predictions then initiate actions that structure our worlds and alter the very things we need to engage and predict. Clark takes us on a journey in discovering the circular causal flows and the self-structuring of the environment that define "the predictive brain." What emerges is a bold, new, cutting-edge vision that reveals the brain as our driving force in the daily surf through the waves of sensory stimulation.
424 pages, Paperback
First published September 15, 2015
177 people are currently reading
2950 people want to read
About the author
Andy Clark
22 books188 followersLibrarian Note: There is more than one author in the GoodReads database with this name. See this thread for more information.
Ratings & Reviews
Friends & Following
Create a free account to discover what your friends think of this book!
Community Reviews
Displaying 1 - 30 of 53 reviews
October 3, 2017
Surfing Uncertainty is contending with the so-called “sense-think-act” model in which a brain passively receives sensory stimuli to incorporate into a model of the world in order to whittle down to a best course of action by first running various mental simulations.
Clark’s alternative “predictive processing” model accounts for a fair degree of regularity in the world—the fact that few situations constitute entirely new experience—presenting a brain which wires itself (in a Bayesian scheme) according to these probabilities to anticipate incoming stimulus rather than simply waiting for its arrival, giving the predictive brain a competitive head start on the one that’d prefer to have all the facts in and tallied first.
So it’s predicting stimulus, not sensing the world per se, that’s the brain’s job, which, as Clark notes throughout, makes for a restless brain. The surfed uncertain interface is where top-down prediction of stimuli—exteroceptive, interoceptive, and proprioceptive—meet the actual stimuli signal. Whatever’s unsuccessfully anticipated by the top-down mechanism migrates up the chain as prediction error, making the predictive brain more computationally efficient since (as compared with the non-predictive brain) only the prediction error—and not the entire stimulus—carries new information.
Another key feature of the predictive processing model is that there’s some estimation of the quality of the signal—how much noise accompanies it. The predictive brain will realize that the precision of the visual signal (and hence the prediction errors it generates) on a foggy day is not as trustworthy as on a clear day. And then the role of attention is to increase the gain.
Prediction and stimuli never completely cancel out. The error (information) flowing up-chain never settles all accounts, and Clark goes to great pains to convey that the predictive model is action-oriented. We predict the world in order to take action and we take action (sampling the world in new ways) in order to enhance prediction. Hence the brain’s restlessness. What we’re trying to do is merely retain a “grip” on the world—or more precisely “a partial grip upon a number of competing affordances: an ‘affordance competition’ that is plausibly resolved only as and when action requires.” Whichever affordance model wins out moment by moment is that which results in the least uncertainty. The affordance competition may sound similar to the sense-think-act model where the brain is running a number of simulations on which to act, but in the predictive model this is all happening in a much more real-time fashion, a way that’s much more integrated with the world, and in a way that integrates rather than segregates planning and action centers in the brain.
Building off of this, Clark’s account of movement is interesting and rather surprising if you’ve never previously come across anything like it. Whereas the goal of perception is predicting input, movement works in the PP model by predicting an outcome that has not yet occurred. If the affordance competition settles on some path of action (associated with the least uncertainty) that requires physical movement (lifting your arm, for example), this constitutes a prediction, and you execute the movement by cancelling out the prediction error (the visuoproprioceptive difference) between where your arm currently rests and where the winning affordance predicts it to be. “In this way, predictions of the unfolding proprioceptive patterns that would be associated with the performance of some action actually bring that action about.”
Extending further:
Schizophrenia provides a nice case study in the book of what can happen when the predictive mechanism goes awry. For example, tickling tickles because the stimulation is erratic and unpredictable. The neurotypical population cannot tickle themselves, the proposed explanation being that when the tickled is also the tickler they’re in a position to predict the stimulation. Schizophrenics, though, are more likely to be able to successfully self-tickle. It might be presumed that the problem in the predictive mechanism in schizophrenics is that the top down prediction is in overdrive, permitting too little of the signal to travel up the chain, but Clark counterintuitively gives the reverse explanation:
The hollow face illusion will help demonstrate the theories more concretely. In the illusion a three-dimensional face which is actually concave appears to be convex because (as the PP model goes) our strong expectation of convexity overrides the actual signal of concavity. Schizophrenics, on the other hand, are more likely to perceive the actual concavity.
Although it’s not immediately clear how predictive processing relates to Clark’s previous work on embodied and extended cognition, Surfing Uncertainty builds nicely from it. But a word of warning to potentially more casual readers: Surfing Uncertainty is written for academics not a popular audience. Here’s a nice video of Clark giving a lecture on the book. It’s a good outline of the book’s ideas, and if it satisfies your curiosity you can probably leave it at that knowing that reading the book is mainly going to fill in the academic intricacies without going much further expanding it. If you’re looking to read one of Clark’s books that’s written with a wider audience in mind look into Natural Born Cyborgs .
Clark’s alternative “predictive processing” model accounts for a fair degree of regularity in the world—the fact that few situations constitute entirely new experience—presenting a brain which wires itself (in a Bayesian scheme) according to these probabilities to anticipate incoming stimulus rather than simply waiting for its arrival, giving the predictive brain a competitive head start on the one that’d prefer to have all the facts in and tallied first.
So it’s predicting stimulus, not sensing the world per se, that’s the brain’s job, which, as Clark notes throughout, makes for a restless brain. The surfed uncertain interface is where top-down prediction of stimuli—exteroceptive, interoceptive, and proprioceptive—meet the actual stimuli signal. Whatever’s unsuccessfully anticipated by the top-down mechanism migrates up the chain as prediction error, making the predictive brain more computationally efficient since (as compared with the non-predictive brain) only the prediction error—and not the entire stimulus—carries new information.
Another key feature of the predictive processing model is that there’s some estimation of the quality of the signal—how much noise accompanies it. The predictive brain will realize that the precision of the visual signal (and hence the prediction errors it generates) on a foggy day is not as trustworthy as on a clear day. And then the role of attention is to increase the gain.
Prediction and stimuli never completely cancel out. The error (information) flowing up-chain never settles all accounts, and Clark goes to great pains to convey that the predictive model is action-oriented. We predict the world in order to take action and we take action (sampling the world in new ways) in order to enhance prediction. Hence the brain’s restlessness. What we’re trying to do is merely retain a “grip” on the world—or more precisely “a partial grip upon a number of competing affordances: an ‘affordance competition’ that is plausibly resolved only as and when action requires.” Whichever affordance model wins out moment by moment is that which results in the least uncertainty. The affordance competition may sound similar to the sense-think-act model where the brain is running a number of simulations on which to act, but in the predictive model this is all happening in a much more real-time fashion, a way that’s much more integrated with the world, and in a way that integrates rather than segregates planning and action centers in the brain.
Building off of this, Clark’s account of movement is interesting and rather surprising if you’ve never previously come across anything like it. Whereas the goal of perception is predicting input, movement works in the PP model by predicting an outcome that has not yet occurred. If the affordance competition settles on some path of action (associated with the least uncertainty) that requires physical movement (lifting your arm, for example), this constitutes a prediction, and you execute the movement by cancelling out the prediction error (the visuoproprioceptive difference) between where your arm currently rests and where the winning affordance predicts it to be. “In this way, predictions of the unfolding proprioceptive patterns that would be associated with the performance of some action actually bring that action about.”
Extending further:
[T]he agent by action calls forth the very world that she knows…We sample the scene in ways that reflect and seek to confirm the grip upon the world that structured the sampling. This is a process that only the ‘fit’ hypotheses (assuming that is understood in a suitably action-oriented manner) survive. At longer timescales…this is the process by which we build designer environments that install new predictions that determine how we behave (how we sample that very environment). We thus build worlds that build minds that expect to act in those kinds of worlds.
Schizophrenia provides a nice case study in the book of what can happen when the predictive mechanism goes awry. For example, tickling tickles because the stimulation is erratic and unpredictable. The neurotypical population cannot tickle themselves, the proposed explanation being that when the tickled is also the tickler they’re in a position to predict the stimulation. Schizophrenics, though, are more likely to be able to successfully self-tickle. It might be presumed that the problem in the predictive mechanism in schizophrenics is that the top down prediction is in overdrive, permitting too little of the signal to travel up the chain, but Clark counterintuitively gives the reverse explanation:
Thus suppose that schizophrenia…actually involves a weakening…of the influence of prior expectations relative to the current sensory evidence. This may strike the reader as odd. Surely, I hear you say, the opposite must be the case, for these subjects appear to allow bizarre high-level beliefs to trump the evidence of their senses! It seems increasingly possible, however, that the arrows of causality move in the other direction. A weakened influence of prior expectations relative to the sensory input may result…in anomalous sensory experiences in which (for example) self-generated action appears (to the agent) to have been externally caused. This in turn may lead to the formation of increasingly strange higher level theories and explanations.
The hollow face illusion will help demonstrate the theories more concretely. In the illusion a three-dimensional face which is actually concave appears to be convex because (as the PP model goes) our strong expectation of convexity overrides the actual signal of concavity. Schizophrenics, on the other hand, are more likely to perceive the actual concavity.
Although it’s not immediately clear how predictive processing relates to Clark’s previous work on embodied and extended cognition, Surfing Uncertainty builds nicely from it. But a word of warning to potentially more casual readers: Surfing Uncertainty is written for academics not a popular audience. Here’s a nice video of Clark giving a lecture on the book. It’s a good outline of the book’s ideas, and if it satisfies your curiosity you can probably leave it at that knowing that reading the book is mainly going to fill in the academic intricacies without going much further expanding it. If you’re looking to read one of Clark’s books that’s written with a wider audience in mind look into Natural Born Cyborgs .
July 31, 2016
A quote that seems to sum up the books main point.
> We alter our predictions to fit the world, and alter the world to fit our predictions.
Main/interesting points
- We are prediction machines, always trying to predict the sensory information we are about to receive. We generate these predictions using a generative model. At the highest level we model hypotheses/causes/hidden variables and use these to generate predictions.
- Errors in prediction cause us to rethink our hypotheses at multiple different levels (maybe the location of an expected stimuli, or something higher level like why it was caused).
- Top level predictions have levels of uncertainty, thus when combined with low level information we have level of uncertainty in our prediction errors. This effects how prediction errors are propagated.
- Prediction errors are mediated by an attention mechanism, it tries to find some balance between top and bottom level information. This means we can attenuate, or amplify prediction errors.
- Action is another mechanism for reducing prediction error (the first being changing hypotheses). e.g. We predict that we will sense our hand move, it isnt, move your hand.
Interesting book.
- Would have liked more details on how it actually works, but I guess that is still being researched.
- A reasonably dense writing style (maybe just the topic). He uses big words and complex sentences...
- Still not not convinced by anything bayesian. (intuitively it is nice, but computationally it is generally intractable. so either brains have found a nice way of doing it, or they dont do it... but, I dont understand bayesian stats particularly well).
> We alter our predictions to fit the world, and alter the world to fit our predictions.
Main/interesting points
- We are prediction machines, always trying to predict the sensory information we are about to receive. We generate these predictions using a generative model. At the highest level we model hypotheses/causes/hidden variables and use these to generate predictions.
- Errors in prediction cause us to rethink our hypotheses at multiple different levels (maybe the location of an expected stimuli, or something higher level like why it was caused).
- Top level predictions have levels of uncertainty, thus when combined with low level information we have level of uncertainty in our prediction errors. This effects how prediction errors are propagated.
- Prediction errors are mediated by an attention mechanism, it tries to find some balance between top and bottom level information. This means we can attenuate, or amplify prediction errors.
- Action is another mechanism for reducing prediction error (the first being changing hypotheses). e.g. We predict that we will sense our hand move, it isnt, move your hand.
Interesting book.
- Would have liked more details on how it actually works, but I guess that is still being researched.
- A reasonably dense writing style (maybe just the topic). He uses big words and complex sentences...
- Still not not convinced by anything bayesian. (intuitively it is nice, but computationally it is generally intractable. so either brains have found a nice way of doing it, or they dont do it... but, I dont understand bayesian stats particularly well).
December 27, 2019
After reading a few Slate Star Codex articles, I got quite interested in predictive processing (PP) and the free energy principle. I read a paper by Friston a while back and came away thinking "this is way more abstract and notationally obfuscated than it needs to be. When I try to mentally fill in the gaps, I can eventually sorta see how these ideas could actually be applied. But I'm having to do 95% of the work to reach that point, so it's not clear if the paper actually contains anything more suggestive equations and buzz-words." And so, I thought I'd give Surfing Uncertainty a go. I mean it's written by a philosopher, so it can't be too mathematically obscure, right? And on the back cover, it says that SU "surveys recent developments in theoretical neurobiology and makes them accessible for general readers and students alike". And Scott Alexander liked it. So surely after reading a book-length treatment of PP I will appreciate just how Good and Useful the theory is. Right?
Not so. Here's the formula that Clark uses to write sections:
1. Give the section a cute name like "Whodunnit?" or "The Unexpected Elephant" to lure the reader in.
2. Spout a bunch of buzz-word gobbledygook about enactive Bayesian predictive hierarchies (cite Friston et al).
3. Talk about some experiment, without actually explaining how it relates to PP (cite al et Friston).
3. Exclaim how elegant and empirically successful everything is (cite Friston's student or collaborator).
4. Draw a diagram in Microsoft Word with lots of arrows and buzzwords (cite Friston's uncle/mailman/dog).
I got up to page 146 before giving up in despair.
To be fair, I do think I have a somewhat better idea of what PP is supposed to be about at a pop-sci level now. Predictions go top-down, errors come back bottom-up. Just like the back-propagation algorithm. Action commands go top-down in the same way as predictions do, and that's why efferent and afferent neurons have the same structure. Also, everything is encoding uncertainties and Bayes' Rule is used at every level.
But how is any of this actually implemented in the brain? Does a higher rate of firing mean lower uncertainty in a neuron? Is it the same neurons which propagate predictions down as errors up, or are there different neurons for these two tasks? Where does the prior distribution at the top of the hierarchy come from? How does the brain do actual computations? Are there any proofs of concept of simplified models of the brain doing something interesting? Are there even any detailed descriptions of what a proof of concept might look like?
All of which is to say that if this is "Friston made accessible to prove that the whole Free-Energy/Predictive-Processing/Bayesian-Brain thing isn't a sham" then the field now looks much more like a sham to me. But there is hope! It turns out that Geoffry Hinton (one of the arch-Gods in the creation myths of machine learning) originated much of free energy and predictive processing stuff. So maybe this area is one of Hinton's less significant ideas going rogue? Or maybe there is a kernel of sense to this nonsense? My plan now is to follow up some of the Hinton citations, and I will decide whether or not to finish this book (and/or dig deeper into the field) based on that.
Anyway, here are my notes:
* Seminal proof of concept for predictive coding: Rao and Ballard 1999
* Hinton predictive coding MNIST: Hinton 2007a, Hinton and Nair 2006, Hinton and Salakhutdinov 2006
* Binocular rivalry: present both eyes with different images (such as with 3D glassses) and the person will alternate in which on they see
* The FFA (part of the brain) was thought to be involved in face feature detection. but if the brain was primed to see a face (by showing a rectangle for milliseconds before showing a face to make an unconscious association) then FFA activity was indistinguishable whether or not a face was actually seen, indicating it is more of a "face predictor" than a "face detector"
* Attention increases upstream flow of info (error propagation)
* Greater uncertainty of prediction weakens upstream flow of info
* People must be predicting as they perceive because sports peoples eyes move to where the ball will be
* Similar with sandwich making (Hayhoe 2003) we keep our eyes just ahead of where the knife is cutting
* (Supposedly) our priors about perception which lead to the cornsweet illusion are provably "Bayes optimal" (Brown and Friston 2012, bottom of page 85). This sounds like a crazy claim to me. How can you prove anything the brain does is Bayes optimal? We have very little idea of how the brain works!
* Visual word form area (VWFA) lights up when looking at letter strings (or when blind people read braile)
* Handwriting has invariants across left/right hand, even feet (p. 88)
* Not hearing expected third beep leads to an imagined start of the beep which abruptly trails off: indicating top-down prediction
* A song you know sounds clearer on a static-y radio than one you don't know because you are able to predict the stimulus and filter out the noise
* Dreaming is about simplifying the predictive model
* Maybe the reason that people will mutually push on each others fingers harder and harder when trying to match forces is the same reason you cannot tickle yourself: you automatically subtract the expected result of your own actions from your perceptions
* What is efference copy and corollary discharge?
Not so. Here's the formula that Clark uses to write sections:
1. Give the section a cute name like "Whodunnit?" or "The Unexpected Elephant" to lure the reader in.
2. Spout a bunch of buzz-word gobbledygook about enactive Bayesian predictive hierarchies (cite Friston et al).
3. Talk about some experiment, without actually explaining how it relates to PP (cite al et Friston).
3. Exclaim how elegant and empirically successful everything is (cite Friston's student or collaborator).
4. Draw a diagram in Microsoft Word with lots of arrows and buzzwords (cite Friston's uncle/mailman/dog).
I got up to page 146 before giving up in despair.
To be fair, I do think I have a somewhat better idea of what PP is supposed to be about at a pop-sci level now. Predictions go top-down, errors come back bottom-up. Just like the back-propagation algorithm. Action commands go top-down in the same way as predictions do, and that's why efferent and afferent neurons have the same structure. Also, everything is encoding uncertainties and Bayes' Rule is used at every level.
But how is any of this actually implemented in the brain? Does a higher rate of firing mean lower uncertainty in a neuron? Is it the same neurons which propagate predictions down as errors up, or are there different neurons for these two tasks? Where does the prior distribution at the top of the hierarchy come from? How does the brain do actual computations? Are there any proofs of concept of simplified models of the brain doing something interesting? Are there even any detailed descriptions of what a proof of concept might look like?
All of which is to say that if this is "Friston made accessible to prove that the whole Free-Energy/Predictive-Processing/Bayesian-Brain thing isn't a sham" then the field now looks much more like a sham to me. But there is hope! It turns out that Geoffry Hinton (one of the arch-Gods in the creation myths of machine learning) originated much of free energy and predictive processing stuff. So maybe this area is one of Hinton's less significant ideas going rogue? Or maybe there is a kernel of sense to this nonsense? My plan now is to follow up some of the Hinton citations, and I will decide whether or not to finish this book (and/or dig deeper into the field) based on that.
Anyway, here are my notes:
* Seminal proof of concept for predictive coding: Rao and Ballard 1999
* Hinton predictive coding MNIST: Hinton 2007a, Hinton and Nair 2006, Hinton and Salakhutdinov 2006
* Binocular rivalry: present both eyes with different images (such as with 3D glassses) and the person will alternate in which on they see
* The FFA (part of the brain) was thought to be involved in face feature detection. but if the brain was primed to see a face (by showing a rectangle for milliseconds before showing a face to make an unconscious association) then FFA activity was indistinguishable whether or not a face was actually seen, indicating it is more of a "face predictor" than a "face detector"
* Attention increases upstream flow of info (error propagation)
* Greater uncertainty of prediction weakens upstream flow of info
* People must be predicting as they perceive because sports peoples eyes move to where the ball will be
* Similar with sandwich making (Hayhoe 2003) we keep our eyes just ahead of where the knife is cutting
* (Supposedly) our priors about perception which lead to the cornsweet illusion are provably "Bayes optimal" (Brown and Friston 2012, bottom of page 85). This sounds like a crazy claim to me. How can you prove anything the brain does is Bayes optimal? We have very little idea of how the brain works!
* Visual word form area (VWFA) lights up when looking at letter strings (or when blind people read braile)
* Handwriting has invariants across left/right hand, even feet (p. 88)
* Not hearing expected third beep leads to an imagined start of the beep which abruptly trails off: indicating top-down prediction
* A song you know sounds clearer on a static-y radio than one you don't know because you are able to predict the stimulus and filter out the noise
* Dreaming is about simplifying the predictive model
* Maybe the reason that people will mutually push on each others fingers harder and harder when trying to match forces is the same reason you cannot tickle yourself: you automatically subtract the expected result of your own actions from your perceptions
* What is efference copy and corollary discharge?
April 16, 2020
Invaluable and insightful theoretical explanation; at the same time, the writing tends to repeat itself and keep talking about how beautiful the brain is according to this theory. Was definitely worth reading for the content, but the style felt like it could have used an editor; frequently had to force myself to keep reading.
June 6, 2017
Very interesting topic and well researched book, but too repetitive and difficult to read for me. I wish it was shorter and more to-the-point.
May 24, 2018
One of the most fascinating books I've ever read, and one of the most difficult books to read I've ever experienced. It reads like a 300 page long research survey paper. It was worth it, though. I'm sure I'll be thinking about it for the rest of my life -- or, at least, unless it all turns out to be wrong!
Somehow, this book manages to weave together how it all works: everything from perception, to learning, to action, to imagination and dreaming, to mental disorders all the way to creativity and how we use and build our world to be temporary brain space. And all with a single, reasonably simple theory. It explains so much with so little. And it makes me want to dig into some of the many, many, many, many, many references to learn more. Even if they're all as initially impenetrable as this.
Somehow, this book manages to weave together how it all works: everything from perception, to learning, to action, to imagination and dreaming, to mental disorders all the way to creativity and how we use and build our world to be temporary brain space. And all with a single, reasonably simple theory. It explains so much with so little. And it makes me want to dig into some of the many, many, many, many, many references to learn more. Even if they're all as initially impenetrable as this.
July 2, 2021
I found this book after reading "7 and a half lessons about the brain" by Lisa Feldman Barrett. I've always been interested in the brain and I wanted to read something more in-depth.
Clark's aim is ambitious. He is trying to offer a model of the brain that is able to explain perception, action, cognition, emotions and more from a purely materialistic worldview. His attempt at doing this is by envisioning our brain as a "prediction machine".
Others will be able to give a better summary of the books content and an evaluation of the science. I am in no position to judge that. What I can say is that this is a very interesting book. The idea and concepts he introduces is intuitive and fascinating. I would give this book 4 or 5 stars based on the content. The reason is only gets 3 is because of the writing. Others have already said it, but it cannot be overstated how inaccessible this book is. Long, complicated sentences and repetitions is something you just have to get used to.
Clark's aim is ambitious. He is trying to offer a model of the brain that is able to explain perception, action, cognition, emotions and more from a purely materialistic worldview. His attempt at doing this is by envisioning our brain as a "prediction machine".
Others will be able to give a better summary of the books content and an evaluation of the science. I am in no position to judge that. What I can say is that this is a very interesting book. The idea and concepts he introduces is intuitive and fascinating. I would give this book 4 or 5 stars based on the content. The reason is only gets 3 is because of the writing. Others have already said it, but it cannot be overstated how inaccessible this book is. Long, complicated sentences and repetitions is something you just have to get used to.
November 28, 2025
Took me 3 starts over 1.5 years to read this book all the way through.
Predictive processing (the broad idea that the brain is a multi-layered predictor of rolling sensory input) is extremely fascinating, and Clark's case for how it ties lots of observations about the brain together is very compelling.
The reason this is not a 5-star rating is that the book is written in the most overcomplicated prose I have ever read.
Predictive processing (the broad idea that the brain is a multi-layered predictor of rolling sensory input) is extremely fascinating, and Clark's case for how it ties lots of observations about the brain together is very compelling.
The reason this is not a 5-star rating is that the book is written in the most overcomplicated prose I have ever read.
January 22, 2018
It might be my unfamiliarity with the field, but I found this book dense and jargon-heavy. There was definitely some really interesting material in there, but it was extremely slow reading relative to other nonfiction.
March 12, 2018
Surfing Uncertainty describes minds as hierarchies of prediction engines. Most behavior involves interactions between a stream of information that uses low-level sensory data to adjust higher level predictive models of the world, and another stream of data coming from high-level models that guides low-level sensory processes to better guess the most likely interpretations of ambiguous sensory evidence.
Clark calls this a predictive processing (PP) model; others refer to is as predictive coding.
The book is full of good ideas, presented in a style that sapped my curiosity.
Jeff Hawkins has a more eloquent book about PP (On Intelligence), which focuses on how PP might be used to create artificial intelligence. The underwhelming progress of the company Hawkins started to capitalize on these ideas suggests it wasn't the breakthrough that AI researchers were groping for. In contrast, Clark focuses on how PP helps us understand existing minds.
The PP model clearly has some value. The book was a bit more thorough than I wanted at demonstrating that. Since I didn't find that particularly new or surprising, I'll focus most of this review on a few loose threads that the book left dangling. So don't treat this as a summary of the book (see Slate Star Codex if you want that, or if my review is too cryptic to understand), but rather as an exploration of the questions that the book provoked me to think about.
Attention
Clark says a good deal about how attention works. It involves high gain[1] for pathways with the most reliable (in expectation) evidence of relevant prediction errors. "But attention is not, PP suggests, itself a mechanism so much as a dimension of a much more fundamental resource."
That clarifies a modest portion of "attention" for me, but I have little confidence that Clark or I have summarized it well enough that you can get much out of it short of wading through most of the book.
This model explains why meditation is hard: I have to convince myself that evidence about my breathing (or whatever I focus on) will generate good evidence that I made a mistaken prediction. Yet it's fairly hard to make a faulty prediction about how my breathing will function over the next second or two. Each layer in the hierarchy will mostly just observe the expected evidence about breathing, which means there's no need to signal any surprises to other parts of the hierarchy, and only the lowest levels in the hierarchy have any error signal to attend to. So the high-level parts of the mind will by default divert attention toward some unrelated context where it might find some important error signals (Did I turn the stove off? Do I have time to respond to that email? Is Trump causing civilization to collapse?).
Autism, ADHD, Schizophrenia
Clark describes schizophrenia and autism as opposites on a spectrum that ranges from high reliance on high-level priors, low attention to sensory data (schizophrenia), to high attention on sensory data and low reliance on high-level priors (autism).
Clark's description provides a nice simple model which predicts many of the sensory and social peculiarities of autism.
E.g. for conversation in noisy environments, the average person relies a fair amount on high-level guesses about what the speaker might say, whereas an autistic person will focus more heavily on parsing the immediate sounds. PP says the high-level guesses provide important guidance to the lower level processing. Whereas most alternatives to PP assume a one-way flow of information from sensory inputs to high-level modules. The PP explanation follows naturally from Clark's description of the spectrum, whereas the one-way information flow models seem to need an ad-hoc assumption about something being broken in autistic speech processing.
Small initial differences in processing conversations will build up over time as the average person learns more about modeling minds (by feedback about whether high-level guesses are accurate), while the autistic person will specialize more in distinguishing each individual word from the background noise. That seems almost enough to explain the social differences between autistic people and average people. But I suspect this is just one of many differences (another might be distraction by prediction errors from background noise).
But what do I make of the reports that autistic people are more likely to have alexithymia, are less aware of thirst, and possibly have less introspective awareness in general? While emotions are often generated by high-level thoughts, detecting emotions seems more like sensory perception. And the less awareness of thirst thing seems really hard to reconcile with a model that predicts sense data are more salient.
One possibility is that some parts of the autistic brain do detect emotions, thirst, etc., properly, and the atypical results are due to that information not spreading into the global workspace. But that doesn't sound like an explanation that Clark would endorse. Attention and the global workspace aren't quite synonyms for consciousness, but they're close enough that I have some presumption against believing that we pay attention to something without having it in our global workspace. I'd expect Clark to admit that autistic people aren't paying much attention to emotions, thirst, etc.
The other possibilities that I see suggest that something is missing from Clark's "sense data to high-level priors" axis. Maybe it needs to be supplemented with an additional axis?
A better model would imply more attention to sources of evidence that provide frequent surprises.
It seems like any good model of these phenomena should also provide insights about ADHD. Clark doesn't mention ADHD, but he provided me with the right hints for me to find this paper:
This leads me to suspect that Clark's axis describes a schizophrenia - ADHD spectrum, and that we need an additional axis to adequately model autism.
I suggest that axis be based on how strongly a person wants to minimize prediction errors, with autistic people having a high desire to minimize error, while schizophrenics and ADHD people having relatively low desires to minimize error.
These two axes imply that for sensory data which produce frequent surprises, such as speech, shiny objects, or scratchy clothing, both ADHD and autistic people pay relatively high attention to the sensory data. Whereas for relatively predictable sensory inputs, such as thirst or happiness, autistic people are more on the schizophrenic end of Clark's spectrum. That similarity in attentional attitudes toward speech help explain why there's some overlap between ADHD and autism, even though they are opposites in many ways.
I'm a little unclear here whether I'm using "minimize error" to mean minimize the number of prediction errors, or some estimate of the magnitude of those errors. Most likely I mean the number of errors that exceed whatever threshold is used to decide whether to propagate an error to some higher layer in the hierarchy. [2]
One interesting implication is that autistic people should be relatively comfortable in social contexts that involve singing familiar songs. I'm unclear whether anyone has good evidence about this, and I expect most autistic people to be too focused on the social context to observe any effects on comfort levels.
I'm sure that autism, ADHD, and schizophrenia have a good deal of complexity that aren't explained by this model, but the broad outlines seem at least moderately close to being right, with a satisfying degree of simplicity.
Miscellaneous points
Clark describes (in Section 4.8) how the mind's equivalent of a utility function is intertwined with expectations about the world, especially expectations about how our bodies will act, in ways that impair our ability to observe a clear-cut utility function. He strongly hints that any alternative would require more computational resources and/or react more slowly. That's bad news for hopes that we'll have a simple way to understand what our robot overlords want.
The PP model helps reduce some of the confusion in the nature-nurture debate by illustrating how highly abstract "hyperpriors" might be mostly learned (before any concrete instances are learned), yet look like innate knowledge to anyone who doesn't understand the PP model. E.g. we might learn that objects tend to be "cohesive, bounded, and rigid" before learning concepts such as "balls, discs, stuffed toys". We really need someone to simplify this idea enough so that the average person who engages in nature-nurture debates can understand it. But that seems beyond Clark's skill or mine.
The PP model helps clarify why we mostly perceive the world as unitary, rather than perceiving probability density functions. Seeing probability densities would be more helpful if perception were implemented separately from action. But we wouldn't function well if our commands to our muscles reflected a probability density over what word to say next. The intertwining of our perception with those commands pushes us toward selecting a single best interpretation of our sensory input, even when a probability distribution would more accurately reflect what we know.
PP helps explain the concept of choking: our muscle movements are generated to match our predictions about how our body will move. If we devote attention to the discrepancy between that predicted motion and our current body position, that will draw attention away from the lower-level discrepancies that are used to guide our muscles to adjust to match their predicted state. Note that this implies some difference between choking at sports versus choking on an SAT test. PP can probably explain choking on the SAT as a weaker version of choking on fluent movements - the part about high-level error signals interfering with more valuable lower-level error signals still seems to apply.
The book is mildly helpful at explaining some cognitive "illusions". E.g. the size-weight illusion results from people using a concept that's better labeled "throwability" when asked about weight.
Conclusion
The book devotes too much attention to the basic question of whether PP describes the basics of human minds, and not quite enough exploration into how to apply PP to a wide variety of contexts.
I didn't need conclusive evidence that the basic PP model was realistic, because once I understood it, it seemed obvious that it enables faster reaction times than alternative models, and there are strong evolutionary pressures for fast reaction times.
I would rather have read a book that focused more on understanding the implications of the PP model.
Footnotes
[1] - I'm a bit fuzzy on what Clark means by gain here - information gain seems most plausible, but the electronics meaning also seems relevant. This is an area where Clark is paraphrasing notoriously hard-to-understand ideas from Friston without much improving their clarity.
[2] - One part of the book describes a potential criticism of PP in which PP implies we'll want to minimize prediction errors by hiding in a dark room (which temporarily minimizes prediction errors, at the cost of increasing prediction errors when hunger/thirst become important). Autism seems to involve an above-average preference for this dark room strategy. (See also Toward a Predictive Theory of Depression).
P.S. While researching this, I stumbled on a story titled Vaccinations Made My Cat Autistic, which I would rephrase as "Admiral Ticklebelly holds a grudge against a human over being vaccinated".
Clark calls this a predictive processing (PP) model; others refer to is as predictive coding.
The book is full of good ideas, presented in a style that sapped my curiosity.
Jeff Hawkins has a more eloquent book about PP (On Intelligence), which focuses on how PP might be used to create artificial intelligence. The underwhelming progress of the company Hawkins started to capitalize on these ideas suggests it wasn't the breakthrough that AI researchers were groping for. In contrast, Clark focuses on how PP helps us understand existing minds.
The PP model clearly has some value. The book was a bit more thorough than I wanted at demonstrating that. Since I didn't find that particularly new or surprising, I'll focus most of this review on a few loose threads that the book left dangling. So don't treat this as a summary of the book (see Slate Star Codex if you want that, or if my review is too cryptic to understand), but rather as an exploration of the questions that the book provoked me to think about.
Attention
Clark says a good deal about how attention works. It involves high gain[1] for pathways with the most reliable (in expectation) evidence of relevant prediction errors. "But attention is not, PP suggests, itself a mechanism so much as a dimension of a much more fundamental resource."
That clarifies a modest portion of "attention" for me, but I have little confidence that Clark or I have summarized it well enough that you can get much out of it short of wading through most of the book.
This model explains why meditation is hard: I have to convince myself that evidence about my breathing (or whatever I focus on) will generate good evidence that I made a mistaken prediction. Yet it's fairly hard to make a faulty prediction about how my breathing will function over the next second or two. Each layer in the hierarchy will mostly just observe the expected evidence about breathing, which means there's no need to signal any surprises to other parts of the hierarchy, and only the lowest levels in the hierarchy have any error signal to attend to. So the high-level parts of the mind will by default divert attention toward some unrelated context where it might find some important error signals (Did I turn the stove off? Do I have time to respond to that email? Is Trump causing civilization to collapse?).
Autism, ADHD, Schizophrenia
Clark describes schizophrenia and autism as opposites on a spectrum that ranges from high reliance on high-level priors, low attention to sensory data (schizophrenia), to high attention on sensory data and low reliance on high-level priors (autism).
Clark's description provides a nice simple model which predicts many of the sensory and social peculiarities of autism.
E.g. for conversation in noisy environments, the average person relies a fair amount on high-level guesses about what the speaker might say, whereas an autistic person will focus more heavily on parsing the immediate sounds. PP says the high-level guesses provide important guidance to the lower level processing. Whereas most alternatives to PP assume a one-way flow of information from sensory inputs to high-level modules. The PP explanation follows naturally from Clark's description of the spectrum, whereas the one-way information flow models seem to need an ad-hoc assumption about something being broken in autistic speech processing.
Small initial differences in processing conversations will build up over time as the average person learns more about modeling minds (by feedback about whether high-level guesses are accurate), while the autistic person will specialize more in distinguishing each individual word from the background noise. That seems almost enough to explain the social differences between autistic people and average people. But I suspect this is just one of many differences (another might be distraction by prediction errors from background noise).
But what do I make of the reports that autistic people are more likely to have alexithymia, are less aware of thirst, and possibly have less introspective awareness in general? While emotions are often generated by high-level thoughts, detecting emotions seems more like sensory perception. And the less awareness of thirst thing seems really hard to reconcile with a model that predicts sense data are more salient.
One possibility is that some parts of the autistic brain do detect emotions, thirst, etc., properly, and the atypical results are due to that information not spreading into the global workspace. But that doesn't sound like an explanation that Clark would endorse. Attention and the global workspace aren't quite synonyms for consciousness, but they're close enough that I have some presumption against believing that we pay attention to something without having it in our global workspace. I'd expect Clark to admit that autistic people aren't paying much attention to emotions, thirst, etc.
The other possibilities that I see suggest that something is missing from Clark's "sense data to high-level priors" axis. Maybe it needs to be supplemented with an additional axis?
A better model would imply more attention to sources of evidence that provide frequent surprises.
It seems like any good model of these phenomena should also provide insights about ADHD. Clark doesn't mention ADHD, but he provided me with the right hints for me to find this paper:
Based on the predictive coding account, top-down expectation abnormalities could be attributed to a disproportionate reliance (precision) allocated to prior beliefs in ASD and to sensory input in ADHD. ... Specifically, difficulties in generating predictions would increase reliance on novel sensory evidence. Accordingly, ADHD individuals (and contrary to ASD subjects) exhibit higher or even exaggerated neural responses to novel/unexpected stimuli (Gumenyuk et al. 2005) and lower responses to expected cues (Marzinzik et al. 2012). ... Based on predictive coding, our results suggest that ASD individuals could be impaired in their ability to adjust precision if faced with uncertainty due to inflexible expectation (Van de Cruys et al. 2014). In other words, the tendency to inhibit bottom-up influences and the attentional bias toward expected stimuli may trigger difficulties in adjusting precision in changing real-world environments. ... Based on the predictive coding framework, our results suggest that difficulties in top-down expectation in children with ADHD are due to high precision ascribed to novel sensory evidence relative to task instructions.
This leads me to suspect that Clark's axis describes a schizophrenia - ADHD spectrum, and that we need an additional axis to adequately model autism.
I suggest that axis be based on how strongly a person wants to minimize prediction errors, with autistic people having a high desire to minimize error, while schizophrenics and ADHD people having relatively low desires to minimize error.
These two axes imply that for sensory data which produce frequent surprises, such as speech, shiny objects, or scratchy clothing, both ADHD and autistic people pay relatively high attention to the sensory data. Whereas for relatively predictable sensory inputs, such as thirst or happiness, autistic people are more on the schizophrenic end of Clark's spectrum. That similarity in attentional attitudes toward speech help explain why there's some overlap between ADHD and autism, even though they are opposites in many ways.
I'm a little unclear here whether I'm using "minimize error" to mean minimize the number of prediction errors, or some estimate of the magnitude of those errors. Most likely I mean the number of errors that exceed whatever threshold is used to decide whether to propagate an error to some higher layer in the hierarchy. [2]
One interesting implication is that autistic people should be relatively comfortable in social contexts that involve singing familiar songs. I'm unclear whether anyone has good evidence about this, and I expect most autistic people to be too focused on the social context to observe any effects on comfort levels.
I'm sure that autism, ADHD, and schizophrenia have a good deal of complexity that aren't explained by this model, but the broad outlines seem at least moderately close to being right, with a satisfying degree of simplicity.
Miscellaneous points
Clark describes (in Section 4.8) how the mind's equivalent of a utility function is intertwined with expectations about the world, especially expectations about how our bodies will act, in ways that impair our ability to observe a clear-cut utility function. He strongly hints that any alternative would require more computational resources and/or react more slowly. That's bad news for hopes that we'll have a simple way to understand what our robot overlords want.
The PP model helps reduce some of the confusion in the nature-nurture debate by illustrating how highly abstract "hyperpriors" might be mostly learned (before any concrete instances are learned), yet look like innate knowledge to anyone who doesn't understand the PP model. E.g. we might learn that objects tend to be "cohesive, bounded, and rigid" before learning concepts such as "balls, discs, stuffed toys". We really need someone to simplify this idea enough so that the average person who engages in nature-nurture debates can understand it. But that seems beyond Clark's skill or mine.
The PP model helps clarify why we mostly perceive the world as unitary, rather than perceiving probability density functions. Seeing probability densities would be more helpful if perception were implemented separately from action. But we wouldn't function well if our commands to our muscles reflected a probability density over what word to say next. The intertwining of our perception with those commands pushes us toward selecting a single best interpretation of our sensory input, even when a probability distribution would more accurately reflect what we know.
PP helps explain the concept of choking: our muscle movements are generated to match our predictions about how our body will move. If we devote attention to the discrepancy between that predicted motion and our current body position, that will draw attention away from the lower-level discrepancies that are used to guide our muscles to adjust to match their predicted state. Note that this implies some difference between choking at sports versus choking on an SAT test. PP can probably explain choking on the SAT as a weaker version of choking on fluent movements - the part about high-level error signals interfering with more valuable lower-level error signals still seems to apply.
The book is mildly helpful at explaining some cognitive "illusions". E.g. the size-weight illusion results from people using a concept that's better labeled "throwability" when asked about weight.
Conclusion
The book devotes too much attention to the basic question of whether PP describes the basics of human minds, and not quite enough exploration into how to apply PP to a wide variety of contexts.
I didn't need conclusive evidence that the basic PP model was realistic, because once I understood it, it seemed obvious that it enables faster reaction times than alternative models, and there are strong evolutionary pressures for fast reaction times.
I would rather have read a book that focused more on understanding the implications of the PP model.
Footnotes
[1] - I'm a bit fuzzy on what Clark means by gain here - information gain seems most plausible, but the electronics meaning also seems relevant. This is an area where Clark is paraphrasing notoriously hard-to-understand ideas from Friston without much improving their clarity.
[2] - One part of the book describes a potential criticism of PP in which PP implies we'll want to minimize prediction errors by hiding in a dark room (which temporarily minimizes prediction errors, at the cost of increasing prediction errors when hunger/thirst become important). Autism seems to involve an above-average preference for this dark room strategy. (See also Toward a Predictive Theory of Depression).
P.S. While researching this, I stumbled on a story titled Vaccinations Made My Cat Autistic, which I would rephrase as "Admiral Ticklebelly holds a grudge against a human over being vaccinated".
August 10, 2021
Very cool 😎
October 19, 2018
Fun? No. Enlightening? Yes.
I wish this book had been written by a more engaging author. The text felt really repetitive and long-winded. I got about halfway through before I admitted to myself I was too bored to keep panning for gold. If you have the time and patience it's worth a read, but know what you're getting yourself into.
My main takeaway from what I read: Seek and you shall find. The first half of the book was basically just a series of convincing arguments to show that the brain operates more from a top-down model (predictive processing) than a bottom-up one (raw input gets descrambled/interpreted). Seems like an intellectualized version of The Secret.
I wish this book had been written by a more engaging author. The text felt really repetitive and long-winded. I got about halfway through before I admitted to myself I was too bored to keep panning for gold. If you have the time and patience it's worth a read, but know what you're getting yourself into.
My main takeaway from what I read: Seek and you shall find. The first half of the book was basically just a series of convincing arguments to show that the brain operates more from a top-down model (predictive processing) than a bottom-up one (raw input gets descrambled/interpreted). Seems like an intellectualized version of The Secret.
Read
August 31, 2017Inspirational insight into the field of embodied cognition. I love the idea of predictive processing and active participation neural on perception. It presents promising theory of how lower level cognitive processes produce output which feeds conscious thought.
October 23, 2019
This book was frustrating. On one hand, it presents a lot of interesting ideas about how the mind might operate, and the theory, if correct, has important implications for things like psychiatry and AI. On the other hand, the prose is a mess and the book is disorganized. I would love a version of this book edited down to 100 pages.
The thesis of the book is that perception, cognition, and consciousness operate as a top-down generative model-- the implication here is that our conscious world is essentially a high-level "hallucinated" model generated by the mind and we rarely have conscious access to our senses. Most mental processing is therefore a complex interaction between high-level top-down predictions and bottom-up residual differences of our raw sensory information with respect to these predictions. This model provides a strong foundation for explaining some of the idiosyncracies of our cognition as well as various disorders of the mind.
Anyhow, there are nuggets of great insight, but they are few and far between, buried beneath of layers of poorly edited and cumbersome prose. At some point I became so annoyed by the author's weirdly overused verbiage that I started compiling some of his "best hits" into a single, terrible paragraph. If the writing style of the quote below doesn't offend you, then I recommend this book:
"The upshot is, if this story is on track, that PP allows (mostly, save for some sideways connections) top-down generation (Friston, Mattout, & Kilner 2011, p. 150) of precision weighted (which is the inverse of variance) predictions that, as we will see in Chapter 5, bridge (what might be thought as) imagination (see Hinton's wake sleep algorithm, 2001), action (via proprioception, more on that later), and distal observations (simulations) of other (embodied) agents-- Pezzulo (2012) describes this as a 'covert loop'-- into a single (generative, if this story holds) neural cognitive framework (or so I shall argue, see Chapter 3), assuming (for the sake of argument) that the story just rehearsed is correct."
The sad part is that while this isn't a direct quote, I would bet that those who have read the book would not be able to detect otherwise.
The thesis of the book is that perception, cognition, and consciousness operate as a top-down generative model-- the implication here is that our conscious world is essentially a high-level "hallucinated" model generated by the mind and we rarely have conscious access to our senses. Most mental processing is therefore a complex interaction between high-level top-down predictions and bottom-up residual differences of our raw sensory information with respect to these predictions. This model provides a strong foundation for explaining some of the idiosyncracies of our cognition as well as various disorders of the mind.
Anyhow, there are nuggets of great insight, but they are few and far between, buried beneath of layers of poorly edited and cumbersome prose. At some point I became so annoyed by the author's weirdly overused verbiage that I started compiling some of his "best hits" into a single, terrible paragraph. If the writing style of the quote below doesn't offend you, then I recommend this book:
"The upshot is, if this story is on track, that PP allows (mostly, save for some sideways connections) top-down generation (Friston, Mattout, & Kilner 2011, p. 150) of precision weighted (which is the inverse of variance) predictions that, as we will see in Chapter 5, bridge (what might be thought as) imagination (see Hinton's wake sleep algorithm, 2001), action (via proprioception, more on that later), and distal observations (simulations) of other (embodied) agents-- Pezzulo (2012) describes this as a 'covert loop'-- into a single (generative, if this story holds) neural cognitive framework (or so I shall argue, see Chapter 3), assuming (for the sake of argument) that the story just rehearsed is correct."
The sad part is that while this isn't a direct quote, I would bet that those who have read the book would not be able to detect otherwise.
December 27, 2023
It seems brains predict (hallucinate) the world about as much as they actually sense what’s there… wild!
February 26, 2018
I've marked this book as read but I haven't read it. I have no doubt that it's an important book and I followed along for a few chapters but it's just too technical for me. I needed to stop 3 times per page to go and look up some terms that he was using then further consider whether he was using those terms literally or metaphorically. It's a rewarding way to read but it's exhausting so for now I'm taking it off my 'currently reading' list so that it stops nagging at me and blocking me from reading other books.
The actual idea in here in the centre of the book is pretty incredible. Essentially it says that human consciousness is solved. It's all about predictive processing. I don't understand how that happens from computing perspective across different timescales and I don't understand exactly how that computing is implemented in biology but I have slightly more of an understanding than I did when I started and I have quite a lot of faith that those within the field understand.
Consciousness seems like such an important topic that I don't want to leave it to the experts so it annoys me that I don't have the ability to catch up with them. I look forward to the Cliffs Notes version!
The actual idea in here in the centre of the book is pretty incredible. Essentially it says that human consciousness is solved. It's all about predictive processing. I don't understand how that happens from computing perspective across different timescales and I don't understand exactly how that computing is implemented in biology but I have slightly more of an understanding than I did when I started and I have quite a lot of faith that those within the field understand.
Consciousness seems like such an important topic that I don't want to leave it to the experts so it annoys me that I don't have the ability to catch up with them. I look forward to the Cliffs Notes version!
March 19, 2017
Andy Clark wrote a fantastic introduction into the topic of predictive processing. The clear advantage of this account is that it enables to combine findings from all disciplines of cognitive science. But there is still much work to do ...
June 20, 2019
This is not an easy read, at least not for those who are not specialists in the philosophy of mind. The book lays out the model of the predictive mind and gives appealing and convincing arguments for this theory. It is definitely worth the effort.
March 5, 2017
Great introduction and overview of predictive models of how the brain and cognition might work
March 20, 2020
A difficult book that often takes several re-readings of passages to get some understanding.
p.1: Rooted in the dynamics of self-organization, these ‘predictive processing’ (PP) models deliver compelling accounts of perception, action, and imaginative simulation.
p.62: The main finding, unsurprisingly, is one of facilitation: valid cues speed up detection of the target stimulus while targets presented on incongruent trials are perceived more slowly, and with less confidence.
p.120-121: Descending pathways in motor cortex, this traditional picture suggests, should correspond functionally to ascending pathways in the visual cortex. This is not, however, the case…. The explanation, PP suggests, is that the downwards connections are, in both cases, taking care of essentially the same kind of business: the business of predicting sensory stimulation.
p.181: …. we leverage three key properties of the predictive processing framework. The first concerns the probabilistic nature of the representations that support perception and action. The second concerns the computational intimacy of perception, cognition, and action. The third concerns the distinctive forms of circular causal interaction between organism and environment that result. Affordance competition then emerges as a natural consequence of probabilistic action-oriented prediction.
p.207: Once higher level stories take hold, new low-level sensory stimulation may be interpreted falsely.
p.218: This pattern of effects,.... might also underlie the everyday experience of ‘choking’ while engaged in some sport or delicate (but well-practiced) physical activity…. In such cases, the deployment of deliberate attention to the movement seems to interfere with our own capacities to produce it with fluency and ease. The problem may be that attending to the movement increases the precision of current sensory information with a consequent decrease in the influence of the higher level proprioceptive predictions that would otherwise entrain fluid movement.
p.239: Much of that progress, we saw, depends upon a swathe of recent empirically informed conjecture concerning the role of ‘interoceptive inference’ —roughly, the prediction and accommodation of our own internal bodily states. Taken together, and mixed liberally with the rich PP account of prediction, action, and imagination, these deliver a startlingly familiar vision: the vision of a creature whose own bodily needs, condition, and sense of physical presence forms the pivot-point for knowing, active encounters with a structured and inherently meaningful external world. This multilayered texture, in which a world of external causes and opportunities for organism-salient action is presented to a creature in a way constantly intermingled with a grip upon its own bodily condition may lie at the very heart of that ever-elusive, and ever-familiar, beast that we call ‘conscious experience’.
p.281: The most obvious example is reading and writing, a matched pair of cultural practices that seem to have emerged far too recently to be a result of genetic adaptations. The practice of reading is known to cause widespread changes in human neural organization.
p.1: Rooted in the dynamics of self-organization, these ‘predictive processing’ (PP) models deliver compelling accounts of perception, action, and imaginative simulation.
p.62: The main finding, unsurprisingly, is one of facilitation: valid cues speed up detection of the target stimulus while targets presented on incongruent trials are perceived more slowly, and with less confidence.
p.120-121: Descending pathways in motor cortex, this traditional picture suggests, should correspond functionally to ascending pathways in the visual cortex. This is not, however, the case…. The explanation, PP suggests, is that the downwards connections are, in both cases, taking care of essentially the same kind of business: the business of predicting sensory stimulation.
p.181: …. we leverage three key properties of the predictive processing framework. The first concerns the probabilistic nature of the representations that support perception and action. The second concerns the computational intimacy of perception, cognition, and action. The third concerns the distinctive forms of circular causal interaction between organism and environment that result. Affordance competition then emerges as a natural consequence of probabilistic action-oriented prediction.
p.207: Once higher level stories take hold, new low-level sensory stimulation may be interpreted falsely.
p.218: This pattern of effects,.... might also underlie the everyday experience of ‘choking’ while engaged in some sport or delicate (but well-practiced) physical activity…. In such cases, the deployment of deliberate attention to the movement seems to interfere with our own capacities to produce it with fluency and ease. The problem may be that attending to the movement increases the precision of current sensory information with a consequent decrease in the influence of the higher level proprioceptive predictions that would otherwise entrain fluid movement.
p.239: Much of that progress, we saw, depends upon a swathe of recent empirically informed conjecture concerning the role of ‘interoceptive inference’ —roughly, the prediction and accommodation of our own internal bodily states. Taken together, and mixed liberally with the rich PP account of prediction, action, and imagination, these deliver a startlingly familiar vision: the vision of a creature whose own bodily needs, condition, and sense of physical presence forms the pivot-point for knowing, active encounters with a structured and inherently meaningful external world. This multilayered texture, in which a world of external causes and opportunities for organism-salient action is presented to a creature in a way constantly intermingled with a grip upon its own bodily condition may lie at the very heart of that ever-elusive, and ever-familiar, beast that we call ‘conscious experience’.
p.281: The most obvious example is reading and writing, a matched pair of cultural practices that seem to have emerged far too recently to be a result of genetic adaptations. The practice of reading is known to cause widespread changes in human neural organization.
May 21, 2019
How do we know what's going on in the world? How do people come to take action based on that understanding - be it catching a ball, typing a book review, learning long division, or building a hadron collider (of any size)? Andy Clark sets out a compelling story of how all this has been coming together in neurophysiology over the last few years into a model based on a hierarchy of processing elements, from the raw sensations upwards, with feedback from the higher levels back to the lower ones.
This model is called 'predictive processing', and if offers a compelling, unified view of how brains seem to work in the world. Crucially, the hierarchies of processing layers in brains don't just sit there waiting for input, but rather they are always predicting what they expect to see coming up from the lower level, noticing any errors, and passing those errors up to the next level whilst simultaneously passing back down an updated prediction as to what might be going on. This means we actively build and re-build models as we go and as we act, rather than waiting for a complete model to become available and then deciding how to act.
The story told here is compelling, detailed, and fits with our emerging understanding both of the data from animal (including human) experiments and of computational models such as neural networks. It explains all sorts of things like dreams, saccade movements of our eyes, and some of the mechanics of language processing. It's a great story, and I suspect the truth as we eventually uncover it will be more or less along these lines.
Having said all of which, I didn't find this an easy book to read. Andy Clark is a Professor of Logic and Metaphysics at Edinburgh University, and even as a sciency person myself, I found his prose style dry and, well, academic. There were a few typos that I spotted, and a bit of a heavier hand with the editing might have helped. My complaints here vary from his desire to make up new words like "surprisal" where we already have perfectly good ones like "suprise", to mixing between a description of hierarchies as sometimes going from front to back, and sometimes going from low to high - occasionally using both in the same sentence. It meant I could only read a few pages at a time before I lost the thread - although I would say there were helpful summary sections at the beginning and end of each chapter, which helped me keep my grip on things, more or less. My favourite of these was for Chapter 8, The Lazy Predictive Brain, which opened up talking about a famous 1989 paper on robotics called "Fast, Cheap, and Out of Control". Which always sounds like a fun Saturday night to me! (or perhaps that's just my brain hierarchy switching from up-down to front-back...)
The book is dense with footnotes and references, which is great if you're another person in this field, but less useful if you're a more general reader. So although this was a very interesting book in many ways, and I suspect really really good if you're in this field, I'm only going to give it 3 1/2 stars for myself, rounded down to three for prose density.
This model is called 'predictive processing', and if offers a compelling, unified view of how brains seem to work in the world. Crucially, the hierarchies of processing layers in brains don't just sit there waiting for input, but rather they are always predicting what they expect to see coming up from the lower level, noticing any errors, and passing those errors up to the next level whilst simultaneously passing back down an updated prediction as to what might be going on. This means we actively build and re-build models as we go and as we act, rather than waiting for a complete model to become available and then deciding how to act.
The story told here is compelling, detailed, and fits with our emerging understanding both of the data from animal (including human) experiments and of computational models such as neural networks. It explains all sorts of things like dreams, saccade movements of our eyes, and some of the mechanics of language processing. It's a great story, and I suspect the truth as we eventually uncover it will be more or less along these lines.
Having said all of which, I didn't find this an easy book to read. Andy Clark is a Professor of Logic and Metaphysics at Edinburgh University, and even as a sciency person myself, I found his prose style dry and, well, academic. There were a few typos that I spotted, and a bit of a heavier hand with the editing might have helped. My complaints here vary from his desire to make up new words like "surprisal" where we already have perfectly good ones like "suprise", to mixing between a description of hierarchies as sometimes going from front to back, and sometimes going from low to high - occasionally using both in the same sentence. It meant I could only read a few pages at a time before I lost the thread - although I would say there were helpful summary sections at the beginning and end of each chapter, which helped me keep my grip on things, more or less. My favourite of these was for Chapter 8, The Lazy Predictive Brain, which opened up talking about a famous 1989 paper on robotics called "Fast, Cheap, and Out of Control". Which always sounds like a fun Saturday night to me! (or perhaps that's just my brain hierarchy switching from up-down to front-back...)
The book is dense with footnotes and references, which is great if you're another person in this field, but less useful if you're a more general reader. So although this was a very interesting book in many ways, and I suspect really really good if you're in this field, I'm only going to give it 3 1/2 stars for myself, rounded down to three for prose density.
April 6, 2025
When you perceive x, your brain is simultaneously making predictions about what x might be, and trying to reconstruct whatever it predicts it to (be likely to) be. Then the predictions are weighted, e.g. in function of both exteroceptive and interoceptive data, so it's possible, e.g., to attenuate specifically self-generated sensory stimuli => you can't self-tickle, except when that goes wrong, and you get schizophrenia.
This means there's a *lot* of cognitive permeability: vision is not uninformed by belief etc., like classical computationalism (cf. Fodor) thought. You can 'sense' things that are not there if you are, say, primed to 'sense' them. What happens in sensory illusions like binocular rivalry or the rubber-hand illusion is a routine that is perfectly adaptive in the wild makes wrong predictions.
We perceive a world that is always-already affordance-laden, presented as relevant for action, and we act on that world, i.e. we manipulate the sensory stimulation that we get (this is part of what acting is). We perceive the world that we made. Perception and action are caught in this feedback loop. That's culture.
It is useful to oppose a reconstructive to a selective or restrictive theory of ‘representation’: the reconstructive theory is one where what the mind does is build a model of what is going on and use that as a tool for planning, detached reasoning, and so on, while the selective theory is one where what the mind does is merely, well, select the (action-)relevant aspects of the environment to enable behaviour (p. 190). It really is the frame problem. “Instead of seeing perception as the control of action, it becomes fruitful to think of action as the control of perception (Powers 1973, Powers et al., 2011). Thus (re)-conceived, the problem becomes ‘not… choosing the right response in light of a given stimulus but… choosing the right stimulus in light of a given goal’ (Anderson, 2014, p. 182-3).” (191) Thus prioritizing action over description in cognition can be understood as anti-descriptivism in a new key. The affordance-ladenness of our ‘representations’ gives some credence to the idea that what perception gives access to is a world that is, in a weak sense (alien to the generality constraint), always-already ‘conceptualized’, or, less committingly, a human world (196). Clark stresses the idea that there is no incompatibility between the idea that we have (direct) access to the external world, and the idea that this world is a human world.
It's a good book, very dense at times, very shallow, and borderline jargony, at some other times. I would benefit from a second read, probably after I've read Hohwy's The Predictive Mind.
This means there's a *lot* of cognitive permeability: vision is not uninformed by belief etc., like classical computationalism (cf. Fodor) thought. You can 'sense' things that are not there if you are, say, primed to 'sense' them. What happens in sensory illusions like binocular rivalry or the rubber-hand illusion is a routine that is perfectly adaptive in the wild makes wrong predictions.
We perceive a world that is always-already affordance-laden, presented as relevant for action, and we act on that world, i.e. we manipulate the sensory stimulation that we get (this is part of what acting is). We perceive the world that we made. Perception and action are caught in this feedback loop. That's culture.
It is useful to oppose a reconstructive to a selective or restrictive theory of ‘representation’: the reconstructive theory is one where what the mind does is build a model of what is going on and use that as a tool for planning, detached reasoning, and so on, while the selective theory is one where what the mind does is merely, well, select the (action-)relevant aspects of the environment to enable behaviour (p. 190). It really is the frame problem. “Instead of seeing perception as the control of action, it becomes fruitful to think of action as the control of perception (Powers 1973, Powers et al., 2011). Thus (re)-conceived, the problem becomes ‘not… choosing the right response in light of a given stimulus but… choosing the right stimulus in light of a given goal’ (Anderson, 2014, p. 182-3).” (191) Thus prioritizing action over description in cognition can be understood as anti-descriptivism in a new key. The affordance-ladenness of our ‘representations’ gives some credence to the idea that what perception gives access to is a world that is, in a weak sense (alien to the generality constraint), always-already ‘conceptualized’, or, less committingly, a human world (196). Clark stresses the idea that there is no incompatibility between the idea that we have (direct) access to the external world, and the idea that this world is a human world.
It's a good book, very dense at times, very shallow, and borderline jargony, at some other times. I would benefit from a second read, probably after I've read Hohwy's The Predictive Mind.
June 8, 2025
I was blown away by the intricate tapestry of the theory of predictive processing. Andy Clark did not shy away from the complexities that predictive processing includes. I love that there is a massive list of supporting citations and materials that give me the opportunity to delve deeper into any particular aspects of the theory that interest me. I would suggest the importance of understanding the predictive processing theory, if you are a neuroscientist, a biological psychologist, or someone who is interested in how the brain works. However, as enthralled by the theory as I was, there were several issues I had with this book:
1. A lack of critical arguments against the predictive processing account.
2. The inaccessibility of the book to general readers and students. The physical copy of the book I have has positive reviews displayed on the back cover by Karl Friston, Jakob Hohwy, and Anil Seth, all of whom are experts within the topic of predictive processing. I do not want to see reviews by experts in the field. I want to see reviews by students who have a decent grasp of basic neuroscience, but have not spent decades developing the very theories they are basically reviewing. Friston describes the book as "accessible for general readers and students". I am an experienced student, and this was a very very opaque, dense book. I get that the theory is complicated, but I do not think that Andy Clark introduced concepts clearly and basic enough for the general reader to understand.
3. Similar to the previous point, I do not think the writing style was effective in informing the reader. Clark is a philosopher, and is clearly not afraid to hide behind the complexity of terms and phrases or to use words in a poetic manner. However, a lot of the time, meaning is obfuscated by the seemingly deliberate refusal to simplify incredibly complicated concepts. I consistently had to consult Google and Chat, which clarified things for me. I think slowly building up the terminology and writing style to that used in actual articles, rather than mimicking the style immediately, would have been more effective. I would say the writing element that most ignited my frustrations while reading this book, however, is the Clark's unnecessary use of brackets in almost every other sentence.
Overall, I am glad I read the book, despite its very challenging content, and the writing style grew on me and became more illuminating nearer to the end.
1. A lack of critical arguments against the predictive processing account.
2. The inaccessibility of the book to general readers and students. The physical copy of the book I have has positive reviews displayed on the back cover by Karl Friston, Jakob Hohwy, and Anil Seth, all of whom are experts within the topic of predictive processing. I do not want to see reviews by experts in the field. I want to see reviews by students who have a decent grasp of basic neuroscience, but have not spent decades developing the very theories they are basically reviewing. Friston describes the book as "accessible for general readers and students". I am an experienced student, and this was a very very opaque, dense book. I get that the theory is complicated, but I do not think that Andy Clark introduced concepts clearly and basic enough for the general reader to understand.
3. Similar to the previous point, I do not think the writing style was effective in informing the reader. Clark is a philosopher, and is clearly not afraid to hide behind the complexity of terms and phrases or to use words in a poetic manner. However, a lot of the time, meaning is obfuscated by the seemingly deliberate refusal to simplify incredibly complicated concepts. I consistently had to consult Google and Chat, which clarified things for me. I think slowly building up the terminology and writing style to that used in actual articles, rather than mimicking the style immediately, would have been more effective. I would say the writing element that most ignited my frustrations while reading this book, however, is the Clark's unnecessary use of brackets in almost every other sentence.
Overall, I am glad I read the book, despite its very challenging content, and the writing style grew on me and became more illuminating nearer to the end.
September 1, 2019
At times challenging, at times repetitive, but overall immensely rewarding and fascinating read. I'm not truly immersed enough in cog sci/neurosci/phil/AI to be able to say whether predictive processing (and Friston's free energy principle) are as powerful as they claim to be, but I am inclined to think they are because they make sense of so many different phenomena. Clark covers a bunch of ground here laying out various aspects of the theory, evidence for the theory, and different ways we can use the theory to explain human behavior. Much of the technical detail that depends on model knowledge or neuroscience knowledge was tough to follow, but I tried my best, because I think in this case nailing down the details is crucial to determining the veracity of the broader claims. I made tons of notes reading this; either places where I wanted to challenge Clark or thought he could extend his discussion further. I think he only scratches the surface in terms of the various domains PP theories can help us explain more elegantly. I thought he spent a little too much time focusing on the embodied and extended aspects of the theory, and in their most strict form I think they're overstated. But there are moments where Clark's ability to describe the looping/embodied/active processes that undergird PP are powerful, and I think provide swaths of insight into the nature of life and many philosophical problems. My two biggest questions are wanting to know more about specific neural realization, and wanting to know better how to link Clark's account with something like Carruther's account of consciousness, and how all of that fits in with "LOT" talk. There are a bunch of other questions and areas I want to learn more about as a result of this book, and at least for now, it is my guiding post for thinking about the brain. Exciting start on what is potentially a paradigm forming theory.
April 23, 2018
I remember during the years as a PhD in Cognitive Science how much I enjoyed Clark’s work. The clarity and elegance of his work on extended mind and embodied cognition was always enjoyable. It’s been some years since that time and it may be that I’m rusted re. scientific literature, but I found this book extremely dry. Or maybe it's just me too rusty for this kind of literature.
The idea that our brain is a proactive bayesian predictive engine, rather than reactive sensory->computation->reaction machine is definitely interesting and it provides a general explanation for a wide variety of phenomena.
A special interesting cases of study include understanding schizophrenia and the autism spectrum using Predictive Processing. It also made me wonder if music and artistic novelty in general could be also explained based on PP.
On summary, the biggest problem with this book is that is dry, excessively academic and repetitive. Clark insist on his idea extensively, but I suspect it could be summarized (without much loss of details) in half of its length. Regardless of this, I think it is very interesting for CogSci and neuroscience students.
The idea that our brain is a proactive bayesian predictive engine, rather than reactive sensory->computation->reaction machine is definitely interesting and it provides a general explanation for a wide variety of phenomena.
A special interesting cases of study include understanding schizophrenia and the autism spectrum using Predictive Processing. It also made me wonder if music and artistic novelty in general could be also explained based on PP.
On summary, the biggest problem with this book is that is dry, excessively academic and repetitive. Clark insist on his idea extensively, but I suspect it could be summarized (without much loss of details) in half of its length. Regardless of this, I think it is very interesting for CogSci and neuroscience students.
Read
November 15, 2021Surfing Uncertainty is a dense, challenging book for a non-specialist to read, but the model of the mind that it lays out is so illuminating that I had to (eventually) hack through it to the end. Clark's "predictive processing" model provides a surprising yet elegant explanation for a wide variety of observations about the mind, our senses, our motor actions, and certain disorders; indeed, several disorders and quirks that weren't discussed in this book seem like they'd be explained intuitively by this model. One cannot see the world, or the operation of one's body in that world, quite the same way after grasping this model of the world and the evidence for it.
I'd very much like for someone to write a book that makes this argument accessible to a popular audience. The author points out at the end that there is still much to discover, but in the meantime knowing about this model of the mind (and senses, and motor action) could point a lot of people in more fruitful directions.
I'd very much like for someone to write a book that makes this argument accessible to a popular audience. The author points out at the end that there is still much to discover, but in the meantime knowing about this model of the mind (and senses, and motor action) could point a lot of people in more fruitful directions.
March 14, 2018
An enthusiastic and rigorous roundup of the "predictive processing" view of the nervous system, pulling together conceptual, imaging, psych, and computational/robotics studies to make the case for a multi-layered bi-directional feedback model (no surprise yet) at all levels where the key insight is that bottom-up feedback to higher generative models takes the form of dynamically-context-weighted *prediction error* signals.
Really excellent wide-ranging applications are given from bootstrapping learning from sensations, motor control to feelings to language, to schizophrenia to autism to visual tracking and relying on other humans... there's a lot more in here, jargon heavy but engaging reading within that.
Really excellent wide-ranging applications are given from bootstrapping learning from sensations, motor control to feelings to language, to schizophrenia to autism to visual tracking and relying on other humans... there's a lot more in here, jargon heavy but engaging reading within that.
October 16, 2021
This is a strange book. One one hand, it is probably the most comprehensive and convincing materialist perspective on how the brain works (that I've read), on the other hand it's pretty much unreadable.
While reading I would often ricochet from "wow, that is an amazing insight" to "wtf, do these words even make sense as a sentence?" If you want to get through this book, you need to come at it with good-faith.
It was fun reading this side-by-side some other books that I've been reading that are much more sensitive to mystical and magical ways of thinking. Often the ideas were in direct conflict with one another and it made for some fun cognitive dissonance.
While reading I would often ricochet from "wow, that is an amazing insight" to "wtf, do these words even make sense as a sentence?" If you want to get through this book, you need to come at it with good-faith.
It was fun reading this side-by-side some other books that I've been reading that are much more sensitive to mystical and magical ways of thinking. Often the ideas were in direct conflict with one another and it made for some fun cognitive dissonance.
September 2, 2022
Some interesting musings, though the research cited, more specifically Friston's work, is not the strongest and some of the principles are unfalsifiable. This is where the philosophical tag is needed. It's fun to build a theory on how cognition and the environment might interact with one another, though a grain of salt is needed. What we actually know and what is presented in this book, shows quite a gap. It boils down to the fact that a good chunk of our knowledge about the brain's functioning is still rudimentary, despite what some authors or current trends in "pop psychology" might insinuate. There certainly is not enough to assume the kind of model proposed in this book.
January 4, 2025
The embodied mind describes the concept of a brain which constantly tries to predict the path of the sensory signals it receives from a body that acts in the world. The predictions come from ever-changing models the brain maintains. The goal of the brain is to minimize the prediction error of the sensory signals which can be done in two ways: changing the prediction (applying different precision weighting of signals) or changing the sensory signal (acting in a different way). Action shapes perception and perception shapes action. It is a cycle and nobody knows how or why it started. The brain looks outwards and it is geared towards engaging the world.
Displaying 1 - 30 of 53 reviews