Phenom Cogn Sci DOI 10.1007/s11097-017-9516-0 Review of surfing uncertainty: prediction, action, and the embodied mind, by Andy Clark, Oxford University Press, 2016 Daniel Williams 1 # Springer Science+Business Media B.V. 2017 In the 1940s, the Cambridge psychologist and philosopher Kenneth Craik hypothesised that organisms use their nervous systems to construct and manipulate an internal model of the world, the chief function of which is prediction: BIf the organism carries a Bsmall-scale model^ of external reality and of its own possible actions within its head, it is able to… react to future situations before they arise, utilize the knowledge of past events in dealing with the present and future, and in every way to react in a much fuller, safer, and more competent manner to the emergencies which face it^ Craik 1943, 61). Craik died at the age of 31 before he could develop this idea in any detail. If the central message of Clark (2016) is along the right lines, however, it turns out he was extremely prescient. Brains, Clark argues, are fundamentally Bprediction machines^, restless and proactive organs that self-organize around the overarching imperative to minimize the error in their predictions of the incoming sensory signal. When recapitulated up the hierarchical structure of the neocortex, this process of prediction error minimization is supposed to account for - well, just about everything. Drawing mostly on the fascinating work of Karl Friston and his group of collaborators, Clark advances the minimization of prediction error as the fundamental operating principle of cognition. The upshot is what he calls Bpredictive processing^ (henceforth PP), an ambitious attempt to explain the whole panoply of cognitive phenomena that constitute minds in terms of this single principle and the complex of neuronal, bodily and environmental resources it enslaves.1 1 In fact, Clark seems to vacillate on whether he accepts this radical form of predictive processing (see below). * Daniel Williams dw473@cam.ac.uk 1 Faculty of Philosophy, Trinity Hall, University of Cambridge, Cambridge, UK Williams D. On first reading this can seem absurd. A single principle underlying all neuronal operations? Biological systems seem to exhibit a mishmash of operations and functions, woven together in a way that reflects what happened to be adaptive in their idiosyncratic ancestral environments and the particularities of their developmental context. And prediction? What could the minimization of prediction error have to do with motivation, action, sequential reasoning, language-use, long-term planning, social coordination, and so on? Further, doesn’t this emphasis on brain-bound predictive models as the locus of behavioural control return us to the bad old days of Cartesianism, internalism, indirectness, neurocentrism, and all those other nasty relics of a misguided model of the mind that Clark (1997, 2008) himself has relentlessly tried to undermine? BSurfing Uncertainty^ has two principal aims, which in tandem with one another seek to allay these concerns. The first is to showcase PP’s explanatory power. Clark marshals a huge body of suggestive evidence in its favour, carefully identifies its many theoretical virtues, and points the reader towards a large and growing body of its applications in cognitive neuroscience, computational psychiatry, robotics, and more. The upshot, he suggests, is not a monolithic and implausible Bnew science of the mind^ (10), 2 but a unique meeting point for the integration of insights from a diverse body of superficially competing approaches in recent cognitive science. The second is an extended attempt to persuade the reader that PP doesn’t have the worrying BCartesian^ implications suggested by its commitment to internal models and probabilistic inference - implications happily identified by Hohwy (2013) in his excellent treatment of PP, for example. Instead, the theory Boffers a vision of the brain that dovetails perfectly… with work on the embodied and environmentally situated mind^ (295). Far from Bback-sliding towards an outmoded BCartesian^ view in which the mind is an insulated inner arena^ (300), PP provides Bthe perfect neurocomputational partner for recent work on the embodied mind^ (1), neatly accommodating its emphasis on the role of extra-neural resources in cognitive mechanisms and vindicating that set of buzzwords that characterise the embodied cognition literature: Baction-oriented^, Bpragmatic^, Bradical^, Bdirect^, Baffordance^, and so on. The result is a truly impressive book: extremely ambitious, beautifully written, admirably clear, and -in most cases, at least - compelling. No review could do justice to the sheer range of material that Clark covers. Instead, I offer a skeletal outline of the core structure and contents of the book, before raising a general worry. Chapter 1 introduces PP’s account of perception and learning. Encased within the skull with access only to the evolving streams of energetic activity at the organism’s sensory transducers, brains must learn about the environment and identify its evolving state in real time. How do they accomplish this miraculous feat? According to PP, by harnessing this sensory input as ongoing feedback to internally generated top-down predictions. In an inversion of traditional wisdom and commonsense, the only information that then propagates up through the perceptual systems is prediction error, the discrepancy between the sensory data predicted and received. By striving to minimize this mismatch, the brain both installs and updates the Bgenerative model^ from which 2 All references just with page numbers are from Clark (2016). Review of surfing uncertainty these predictions are issued, such that Bthe very same process^ underlies Bboth learning and online response^ (15). Crucially, this process requires hierarchical representation of the sort familiar from deep learning models in contemporary machine learning, extracting hidden causes (latent variables) at multiple levels of spatial and temporal scale. In practice, this means that the cortical model that underlies perception is decomposed into a hierarchy of levels trafficking in representations at different levels of abstraction, with each level attempting to predict activity at the level below. In this way brains identify the set of interacting worldly causes that best explains their evolving sensory inputs by striving to minimize the hierarchically distributed error signals in their predictions of these inputs. Despite the attractions of this picture, I wish Clark had gone a bit deeper into how it is supposed to work. For example, he repeatedly states (e.g. 24; 173) that hierarchical generative models can answer the worries that many have raised about the ability of non-symbolic computational systems to deal with structure on the grounds that Bprediction-driven learning… tends to separate out interacting distal… causes operating at varying scales of space and time^ (24). But the classicist’s worry (e.g. Fodor and Pylyshyn 1988) is not with the ability of non-symbolic architectures to represent structure as such, but with their ability to flexibly recombine the elements implicated in such structured representations in systematic and productive ways. While Clark does an excellent job conveying the current excitement about multilevel graphical models and statistical learning, it wasn’t obvious to me that the cortically realised hierarchical generative models at the core of PP solve this problem.3 Further, Clark claims that PP implements Bayesian inference but doesn’t explain how this is supposed to work. For example, on the assumption that priors are given by predictions, are likelihoods also explicitly encoded? When does the brain redistribute probabilities over predictive hypotheses (perception) and when does it update the hypothesis space itself (learning)? How do predictions that don’t minimize prediction error on a given occasion get entrenched in the brain’s internal model (as they must if its account of perceptual illusions is to work)? There are some gestures towards Blongterm^ and Bglobal^ (200) prediction error minimization, but a sceptical reader would likely want a bit more. Chapter 2 turns to the role of precision-weighting and attention in PP. Precisionweighting is the process by which the brain evaluates its sensory uncertainty, assigning differential weight to prediction errors throughout the perceptual hierarchy in accordance with their estimated reliability. This Bmetacognitive ploy^ (204) enables the brain to modulate the relative influence of prior expectations and incoming information in a flexible way, facilitating the extraction of signal from noise in the incoming sensory data. Following others (e.g. Hohwy 2013), Clark contends that this process of precision-weighting realises the functional role of attention, and it plays a pivotal role throughout the book. Chapter 3 explores how the PP machinery might be harnessed Boff-line^ for the purposes of simulation, counterfactual reasoning, mental time-travel, and more. Because predictive brains are fundamentally in the business of endogenously constructing virtual versions of the sensory data they receive from the environment, perception and 3 Clark seems to acknowledge this in a footnote (320, fn.9). Williams D. Bsomething functionally akin to imagination^ (94) are attractively co-emergent. In this way, Clark suggests that PP can accommodate recent work that treats putatively Bhigher-cognition^ as grounded in more basic forms of sensorimotor processing. Chapter 4 advances PP’s account of action. Following Friston, Clark argues that action is just a different means of minimizing prediction error: instead of updating topdown predictions to bring them into alignment with the incoming signal (perception), action arises as a means of bringing sensory input into alignment with top-down predictions. Motor control, for example, is a matter of first predicting the proprioceptive sensory inputs implied by a desired action, and then activating classical reflex arcs to minimize the resultant error signals conditioned by the absence of that action. Where do such desires or goals come from? Clark (129) suggests that they are just further, higherlevel predictions. Greater elaboration on this point would have been helpful, however. In the conceptual framework of folk psychology, at least, what one wants and what one expects to do are orthogonal categories that often come apart. With the core account of learning, perception, attention, and action on the table, Clark turns to elaborations and applications of the basic story to other phenomena. Chapter 5 deals (among many other fascinating topics) with how predictive minds can harness their generative machinery for the purposes of social coordination, and how precision-weighting can reconfigure the distribution of cognitive influence within the brain in flexible ways. Chapter 6 considers the Bmind-world relation^ from the perspective of PP, challenging Hohwy’s (2013) view that the predictive mind is a fundamentally Cartesian mind, estranged from its world behind a veil of sensory transduction. Not so, Clark argues: perception is Bnot-indirect perception^ (195), whatever that means. Indeed, Clark (202) finishes this chapter by noting that Bwe should probably not worry too much about the words we use here^. This made me wonder why Clark worries so much about the words he uses throughout the book. BThe predictive brain^, he argues, Bis not an insulated inference engine so much as an action-oriented engagement machine^ (1). Perception is not Bcontrolled hallucination^; hallucination is Buncontrolled perception^ (196). And so on. For those of us not quite sure what the Bmind-world relation^ is supposed to be, however, or how one might empirically adjudicate whether perception is Bdirect^ or not, these excursions often give the impression of aesthetic preference rather than metaphysical discovery - a good thing, perhaps, if one is sceptical of the possibility of metaphysical discovery, but less engaging for readers of a more positivist bent. Chapter 7 turns to the vexed issue of consciousness from the perspective of PP, which Clark addresses both through fascinating models of various psychopathologies and through an exploration of interoceptive prediction in the construction of emotional experience. The final two chapters of the book then focus on embodied cognition and human uniqueness within the context of PP. Chapter 8 argues that the apparent tension between PP and work in embodied cognition is illusory, and that the former can - indeed must - straightforwardly assimilate the core theoretical lessons of the latter. Bearing most of the weight here is Clark’s claim that predictive brains are striving not for veridical representation but to keep Bthe organism within its own window of viability^ (269), an aspiration often better served through Bproductive laziness^ (244) than Bexpensive, representation-heavy strategies^ Review of surfing uncertainty (268). In fact, self-organizing around the minimization of prediction error ensures the best of both worlds, enslaving Bfast, cheap modes of response^ alongside Bmore costly, effortful strategies^ depending on task-relevant context (244). Chapter 9 is a fascinating but heavily speculative answer to a worry one might have with PP: if the theory correctly characterises (at least) the neurocomputational architecture of the mammalian cortex, what accounts for the seemingly novel kinds of cognition exhibited by human beings? In an enthralling gesture towards future research, Clark identifies numerous ways in which predictive brains might be augmented and transformed in the idiosyncratic environment of shared cultural practices and flexible symbol systems characteristic of human life. The book concludes with an identification of residual Bpuzzles and problems^, chief among which is the need to explore Bthe full space of possible prediction-andgenerative-model-based architectures and strategies^ (298). This skeletal overview should give some indication of the ambition of the book: from the beginnings of a brain attempting to learn about the world from its sensory effects, Clark traverses the whole spectrum of psychological phenomena that make up the mind, in each case showcasing how prediction-based mechanisms offer illuminating and often surprising explanations. The upshot is an exhilarating reading experience, and anyone interested in cognitive science - indeed, anyone interested in the mind should buy a copy of the book immediately. Despite this, I came away with a nagging concern: it is not always clear which theory Clark is defending. In Chapter 1, for example, he notes that the book explores Btwo distinct but overlapping stories^: BThe first is a general, and increasingly well-supported, vision of the brain (and especially the neocortex) as fundamentally an inner engine of probabilistic prediction. The other is one specific proposal… describing the possible shape and nature of that core process of multilevel probabilistic prediction^ (27-8). Here and elsewhere, Clark suggests that PP represents Bjust one point in the large and complex space of probabilistic generative-model based approaches^ (298), noting that Bwe should not forget… that there are many possible models in this general vicinity^ (28). This seems disingenuous, however, for two reasons. First, whilst there is increasing evidence and theoretical consensus around the idea that one core function of the neocortex is probabilistic prediction, the idea that prediction or the minimization of prediction error constitutes its exclusive function indeed, that it has a single, overarching function -is not widely shared and dramatically underdetermined by current evidence. That’s not to say that the idea is wrong, of course, and Clark gives it about the best possible defence I could imagine. The point is that this presentation masks PP’s radicalism by trying to subsume it under this broader class of models. Most advocates of PP in the literature justify this radicalism by recourse to Friston’s (2010) Bfree-energy principle^, which views prediction error minimization as a special case of a more fundamental imperative in biological systems to selforganize under conditions tending towards increasing disorder. Clark, however, relegates the free-energy principle to a couple of pages in an appendix, and often writes Williams D. as if PP is theoretically independent of it. Indeed, at points Clark explicitly repudiates it as a grand unified theory of brain function (e.g. 265). Short of some principled justification for the radicalism that the free-energy principle aspires to provide, however, it becomes unclear why Clark focuses so heavily on the peculiarities of the PP account. This leads to a second issue: many of the substantive claims that Clark advances about the mind seem to rely on PP’s peculiarities, and not the Bbroader^ class of models that stress the importance of generative-model-based prediction. Indeed, in some cases the two seem to pull in diametrically opposed directions. To take one example, the idea that Bworld-engaging action^ (168) and the maintenance of the organism within its Bwindow of viability^ (192) constitutes the true function of prediction error minimization is one of the central and recurring themes of the book. And it makes perfect sense if one embraces the free-energy principle. In the Bayesian literature untainted by this principle, however, it is explicit that the function of the perceptual system is to transition from proximal sensory inputs to accurate estimations of environmental states (cf. Rescorla 2013). For this reason, much of Clark’s positive vision of the nature of cognition in the book is emphatically not the upshot of a broad class of probabilisticmodel-based approaches to cognition. I suspect Clark might have two kinds of responses to this worry. The first is that there is sufficient evidence for PP considered independently of any considerations of the free-energy principle to warrant his attention. There are two problems with this, however. First, most of the evidence that Clark identifies is evidence for the role of probabilistic prediction, not for PP as such. Where Clark does focus on specific PP models of various cognitive phenomena, they are typically extremely schematic and speculative - something which Clark repeatedly notes himself (104; 204; 298). Second, much of the excitement of PP comes from its theoretical ambition - its aspiration to provide a Btruly fundamental and deeply unified science of the embodied mind (297). To put it mildly, however, there is not compelling evidence that Bthe whole embodied, active system… self-organizes around the organismically-computable quantity ‘prediction error’^ (294). This idea is vastly underdetermined by current neurophysiological and behavioural evidence. In fact, short of the principled justification for this idea of the sort the free-energy principle aspires to provide, it is difficult to take it seriously. Nevertheless, many of Clark’s substantive claims about the nature of cognition seem to rely on it. Second, Clark might respond that the emphasis on the functional primacy of action and pragmatic coping over veridical representation is not specific to the free-energy principle. Clark (1997, 2008) himself, for example, has advocated an Baction-oriented^ approach to the mind long before that principle became fashionable, and the focus on homeostasis, regulation and control as the functional core of cognition is common to a much broader range of approaches (e.g. Anderson 2014; Cisek 1999). Fair enough. But notice that this confirms my point: what Clark is really advancing is not one among a broader class of generative-model-based predictive approaches to cognition, but rather one such approach interpreted in a broader theoretical context that has nothing to do with them. That’s a perfectly valuable project, and Clark carries it out with a skill, sophistication and erudition unparalleled by anyone else in the philosophy of cognitive science. But it is a different project to the one he sometimes claims to be involved in. Review of surfing uncertainty References Anderson, M. (2014). After phrenology (1st ed.). Cambridge, Massachusetts: The MIT Press. Cisek, P. (1999). Beyond the computer metaphor: Behaviour as interaction. Journal of Consciousness Studies, 6, 125–142. Clark, A. (1997). Being there (1st ed.). Cambridge, Mass: MIT Press. Clark, A. (2008). Supersizing the mind (1st ed.). Oxford: Oxford University Press. Clark, A. (2016). Surfing uncertainty (1st ed.). Oxford: Oxford University Press. Craik, K. (1943). The explanation of behaviour (1st ed.). Cambridge: Cambridge University Press. Fodor, J., & Pylyshyn, Z. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1–2), 3–71. doi:10.1016/0010-0277(88)90031-5. Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. Hohwy, J. (2013). The predictive mind (1st ed.). Oxford University Press. Rescorla, M. (2013). Bayesian perceptual psychology. In M. Matthen (Ed.), The Oxford handbook of philosophy of perception (1st ed.). Oxford: Oxford University Press.