The anticipating brain is not a scientist: the free-energy principle from an ecological-enactive perspective

Second, we have argued that free-energy minimization is best understood in terms of the disattunement of the internal and external dynamics of the animal and its environment. It is the whole organism in regulating its interactions with the environmental landscape of affordances that is minimizing surprisal. We will see in the next section, how Jakob Hohwy argues that minimizing surprisal is the evolutionary function of the brain. Hohwy takes the minimization of surprisal to be a function of the brain in the same sense as the function of the heart is to pump blood. We’ve argued by contrast that minimization of surprisal is done by the whole organism being drawn to act on relevant affordances in ways that result in the reduction of dis-attunement with the environment.

Jakob Hohwy has drawn on the free-energy principle to develop a theory of the predictive brain that is very different from the ecological and enactive one we have just proposed. He interprets the free-energy principle using older ideas from Helmholtz (1860/1962) and Gregory (1980), following up on suggestions that can also be found in Friston’s writings (see for instance Friston and Kiebel 2009). Hohwy rightly points out that the Helmholtzian theory of perception is strongly internalist in a way that challenges enactive and ecological approaches towards the mind. In The Predictive Mind (2013), Hohwy has given one of the most comprehensive and careful expositions of the prediction-error minimization framework (PEM) to date. In a more recent follow-up paper, he has the following to say about what his interpretation of PEM implies for a philosophical account of the mind-world relationship:PEM should make us resist conceptions of this relation on which the mind is in some fundamental way open or porous to the world, or on which it is in some strong sense embodied, extended or enactive. Instead, PEM reveals the mind to be inferentially secluded from the world, it seems to be more neurocentrically skull-bound than embodied or extended, and action itself is more an inferential process on sensory input than an enactive coupling with the environment. (Hohwy 2014, p. 1)Generally in formulations of PEM, brain processes are taken to work according to a unifying principle related to but not identical with the free energy principle. PEM claims this unifying principle is the minimization of prediction-error between the actual inflow of sensory input occurring over multiple temporal and spatial scales, and the flow of sensory input the brain predicts based on its top-down knowledge.

Sensory processing in the brain is not a bottom-up process of feature detection but instead consists of the top-down prediction of sensory signals based on hierarchically organised models of environmental causal regularities. What gets conveyed through bottom-up processing are weighted prediction errors that arise when there is a discrepancy, or “prediction-error” between incoming inputs, and what the brain expects. Sensory processing thus becomes about predicting sensory input. By minimizing prediction-errors over time (given a suitable generative model) the system “will come to infer the hidden state of the environment”.

Hohwy equates PEM with the view of perception as knowledge-driven perceptual inference as advocated by Helmholtz and Gregory. The logic is the following: the brain only has access to its own predictions and to prediction-errors and based on these it has to infer the hypothesis that best explains away current prediction error. Perception then essentially is hypothesis-testing in which the brain seeks to explain away prediction-errors by finding the hypothesis that best explains the available sensory evidence. The ‘unconscious inferences’ that lie between a sensation and a perception are much like the inferences that scientists draw (Gregory 1980; Hatfield 2009). It is the inferred hypothesis about the causes of sensory input that provides our epistemic access to the external world. The brain only has access to the ways its own patterns of neural activity and spike trains flow and alter. The true states of the external environment are hidden from the brain beyond the veil of sensory information. Any epistemic access to the world is therefore, according to Helmholtzians like Hohwy, indirect based on the hypothesis that best explains away current prediction error.

This Helmholtzian line of reasoning assumes however a distinction between sensation and perception that is problematic for ecological and enactive cognitive scientists. The general disagreement between inferential and ecological approaches to perception concerns the richness or sparsity of our sensory contact with the environment. For the Helmholtzian, sensations are impoverished and need to be enriched by top-down knowledge before the perceiver can know what is out there in the world. This view of perception as starting from impoverished and ambiguous sensations does not fit with the richness of the ecological context. For the ecological theorist, the perceptual system is attuned to directly (i.e. non-inferentially) pick up on environmental regularities in the organism’s niche. Within that framework, perception is understood in terms of lower-level phenomena like resonance and attunement, mediated by the organism’s skills.21

As we have argued for above, the animal species and the ecological niche are co-specifying. The niche of the animal is best understood as a landscape of affordances that reflects the perceptual capacities and other abilities of a kind of animal belonging to a particular form of life (Rietveld and Kiverstein 2014). Affordances are more precisely defined as relations between aspects of the material environment and abilities available in a ‘form of life’ (ibid., p. 337).22

The lived perspective of an individual organism reflects its needs, concerns and abilities and determines which affordances are relevant in a particular situation. Hohwy is absolutely correct to pick up on a tension between Helmholtz and conceptions of the mind like our own that stress openness to the world. However, unlike Hohwy we take the free-energy principle to challenge those that would follow Helmholtz.

Hohwy conflates PEM with the free-energy principle. He writes: “Since the sum of prediction error over time is also known as free-energy, PEM is also known as the free-energy principle (Friston and Stephan 2007).” (Hohwy 2014, p. 2) We believe this is confusing, if not simply wrong. There is a difference between the free-energy bound and the free-energy principle. The free-energy bound is the computational trick of reducing an (intractable) integration problem to an optimization problem by introducing a variational density.23 The free-energy principle, on the other hand, is tied to the minimization of surprisal and therefore (by the logic of the free-energy principle) to self-organization and what we call the tendency towards an optimal grip on affordances. As we have shown in Sect. 1, one important consequence of the free-energy principle is that it is only action that minimizes surprisal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{E}$$\end{document}E. Perception, understood as changing the internal dynamics of the organism, will only lower the upper-bound on surprisal\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{E}$$\end{document}E and not change surprisal itself.

For Hohwy, perceptual inference is changing the hypothesis so that it fits the data, whereas active inference is changing the data so as to fit the hypothesis. Both are different strategies for minimizing prediction-error. Hohwy has a particular way of adding action to perceptual inference.24 In perception the brain tests its current hypothesis of the world by predicting the sensory input or data it would expect on the assumption that its current hypothesis is true. It then compares the predicted data with the actual data and goes on to adapt its hypothesis to improve the fit with the actual data. The brain laboriously changes and fine-tunes its hypothesis as experimental results come in. These iterations of hypothesis-testing are what Hohwy has in mind when he talks about perceptual inference. He introduces active inference by appealing further to the analogy between perception and scientific hypothesis testing (Hohwy 2013, p. 77). Scientific hypothesis testing is not just passively recording results of an experiment, but carefully setting up experiments and actively intervening in the chains of causes and effects in order to disentangle the causal web. In action the brain controls movement, sampling the environment with the aim of matching actual sensory input with predicted sensory input. According to this framework, the role of action is to support perceptual inference and makes predictions more reliable (Hohwy 2013, p. 79).

However, we have seen that form the perspective of the free-energy principle, perceptual inference without active inference does not make sense, since only action minimizes surprisal. Let us return to the example of temperature we used in Sect. 1. If you find yourself too hot, one way to reduce your temperature is by taking a cold shower. Action here is doing the work of reducing surprisal\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\mathrm{E}}$$\end{document}E, because of its effect on the animal’s relation to the environment, taking it closer to an optimal bodily condition. Perceptual inference also works in the service of actively improving the organism’s grip on the environment. We’ve followed Friston in characterising this form of inference in terms of change in the animal’s internal (endogenous) dynamics. Perceptual inference can thus be thought of as patterns of action readiness. In the case of the shower, the sensation of being too hot shapes your perspective on the environment in such a way that the cold shower now stands out as inviting or attractive (whereas normally it probably would not). Perceptual and active inference are not two distinct strategies for minimizing prediction-error (each with their own so-called ‘directions of fit’), as Hohwy suggests. Perceptual and active inference are entangled parts of a single process of readying the organism to act in such a way as to improve its grip on the environment. Perception, understood as changing the organism’s internal dynamics reflects the organism’s pattern of action readiness and thus its selective openness to affordances.

The organism is ready to act on relevant affordances so as to improve its grip on the environment thereby reducing surprisal. There is no separation between perception and action because perception on our account just is the organism’s preparing itself to act in ways that reduce surprisal, thereby improving grip on its environment both by getting ready for what is to come and by engaging with affordances offered by it in action. The environmental affordances scaffold the individual’s actions. Action does not have a world-to-mind direction of fit, as Hohwy argues, because in acting the agent has no intention or goal in mind. Instead the agent is drawn to act based on its disattunement with the environment. It is the affordances relevant for reducing this disattunement that drive the actions of the agent.

Both Hohwy and Friston acknowledge the central role for action in the minimization of surprisal (Hohwy 2013, pp. 84–88), but we think they fail to realize the damage it does to their Helmholtzian commitments. As mentioned earlier in this paper, on a Helmholtzian account of the mind, the aim of perceptual inference can be said to be to infer the most likely cause of sensory input, the more objective the better. However, we would argue that a perfect hypothesis that precisely represents the state of the environment is worthless if it does not specify what action minimizes surprisal, or improves grip. Hohwy recognizes this point when he writes as follows: “perceptual inference can make you perceive that you are hurtling towards the bottom of the sea with something heavy around your feet but cannot do anything to change that disturbing sensory input which is fast taking you outside your expected states.” (Hohwy 2013, p. 85) It is only by action that you can do something about this disturbing sensory input by, for instance, untying the heavy object from your feet. However, active inference of the type involved in such a life or death situation, would be ecologically useless if it did not predict adaptive actions that would improve my situation in this life threatening environment. The possibility of cutting the rope with my Swiss army knife in my pocket should become the relevant affordance that drives my actions now.

The upshot of this is that my brain is not, and should not behave like an exemplary scientist. If my brain really is a scientist, then it is heavily invested in ensuring the truth of a particular theory, which is the theory that “I am alive”. This is a fundamental prior belief that drives all action; namely, I exist and I will gather all the evidence at hand to prove it. It will only make predictions whose confirmation is line with this hypothesis. It does not give competing hypotheses a fair chance and is extremely biased in the way it interprets the data. It decides on the outcome of an experiment beforehand (my staying alive) and manipulates the experiment until the desired result is reached. If my brain is a scientist, it is a crooked and fraudulent scientist—but the only sort of scientist that can survive an inconstant and capricious world.25 The hypotheses that the brain in reality is biased in favor of, are hypotheses that predict the animal will tend towards grip on environmental affordances.

The interesting question in active inference is therefore not, as Hohwy claims, how the brain can use available sensory input to accurately reconstruct the hidden state of affairs in the world (Hohwy 2014, p. 1). The interesting question is rather how the space of possible ‘hypotheses’ is always already constrained in such a way as to make the animal improve its grip on the environment.

Hohwy might seem to recognize this point when he discusses the limitations of actions understood as active sampling of sensory input. However, he attempts to offer an account of PEM that abstracts away from questions about how the space of hypothesis is already constrained: “I will not engage these types of questions directly here. I am primarily interested in what the account says about our understanding of the world and our place in it as perceivers and agents.” (Hohwy 2013, p. 88) We, however, think that this is where the free-energy principle starts to get interesting: one cannot understand our place as agents in the world without taking into account self-organization. The free-energy principle with its focus on self-organizing dynamical systems is fundamental to answering these questions about the mind-world relation. The Helmholtzian metaphor of the brain as hypothesis-testing scientist is ill-equipped to capture active inference under the free-energy principle for the simple reason that it triggers the wrong questions. The central question for the organism is not what the state of affairs of the environment is (although this might be relevant for achieving certain forms of grip), but rather which actions will minimize surprisal, where surprisal is taken relative to an encultured, skilled and embodied agent. The organism is always acting on the basis of a basic concern to improve grip, and this biases and informs all of its actions. This basic concern is taken for granted or presupposed by the Helmholtzian.

This presupposition shows up for instance in the way in which active inference is informed by and depends upon the goals and intentions of the agent. For example, suppose I am falling down the stairs. I could perfectly predict myself tripping down the stairs and breaking my neck. Navigating the stairs is seen by the Helmholtzian as the goal of the agent, and relative to that goal active inference works so as to bring about perceptions that take you closer to accomplishing that goal. However, it remains a mystery why the agent has selected the goals that it has in this situation. Implicit in this appeal to goals and intentions is the basic concern of the organism towards improving its grip on its environment. The point per se is not about prediction-error minimization, because my brain can be correctly minimizing prediction-error as I fall down the stairs. Although my sensations might tell me it is very likely that I am tripping down the stairs, my brain needs to treat this as a highly unlikely event and do as much as possible to return to a more likely situation (i.e. balancing on two feet). These considerations are echoed in Friston’s (2011) account of embodied inference:I model myself as embodied in my environment and harvest sensory evidence for that model. If I am what I model, then confirmatory evidence will be available. If I am not, then I will experience things that are incompatible with my (hypothetical) existence. And, after a short period, will cease to exist in my present form (Ibid., p. 117)The Helmholtzian is thus already taking for granted without explaining something which is central to our account, the lived perspective and the concerns of the organism. Rather than simply presupposing goals and intentions, we suggest that the basic concern of the organism is tending towards improving its grip on available affordances in its environment.

Unlike Hohwy (2014) and Clark (2013)26 we follow Friston in taking the free-energy principle to be the biological principle of living systems as a whole, and not only of the brains of living systems. Goals and intentions ought to not be something external to the free energy principle, but rather should be explained by the principle itself. The Helmholtzian account does not in itself provide any explanation of the goals of the organism, and why the organism selects the goals it does. It rather specifies prediction-error minimization in relation to those presupposed goals without offering any explanation of the process by which the goals themselves are formed. By contrast our account understands free-energy minimization in terms of the tendency towards an optimal grip on available affordances. It thereby shows how selective openness to relevant affordances follows from acting according to the free-energy principle. There is therefore no need to presuppose goals and intentions whose origins remain quite mysterious, or for an additional account of goal and intention formation.

Our concept of the tendency towards an optimal grip additionally puts the Bayesian notions of precision and uncertainty in a different perspective.27 The traditional function of uncertainty in Bayesian accounts of perception is to modulate the impact of sensory perturbations (‘prediction-errors’) on the internal dynamics (‘hypothesis’). If the agent has high confidence in its sensory input and low confidence in its current hypothesis, then the sensory prediction-error will greatly shift the hypothesis. In the reversed case (low sensory confidence and high internal confidence), the probability density will be unaltered.

Confidence however cannot be a function of sensory evidence alone taken in isolation from contexts of skilled action and engagement with the world. The central question for the agent and thus for active inference is not to settle on which hypothesis is true, but on what needs to be done. Even as all sensory evidence points to me standing under a shower that is too hot, I need to treat this as an unlikely event and arrange my actions accordingly. Moreover, most of our actions are extremely context-sensitive: the ringing phone solicits answering when I am alone, for example, but not when I am having a conversation. Situations typically offer multiple possibilities for action, and the degree to which each of these possibilities stands out as salient or relevant is due to precision-modulation.28 Crucially, such precision-modulation is structured by an agent’s skills and concern—their acquaintance with a normative socio-cultural practice (e.g. when it’s acceptable to answer your phone); their habits (getting a cup of coffee before starting work); their bodily needs (eating an apple when hungry). While engaged in a conversation a buzzing phone leaves us cold (does not alter internal dynamics, has low precision), while in another context it solicits answering (high precision, impacts internal dynamics). Precision-modulation based on these kinds of factors, shapes the salience and relevance of the field of affordances with which agents engage. We cannot understand our place as agents in the world without taking into account the wider contexts and situations in which skilled action takes place.

Both Hohwy (2014) and Clark (2013) take the brain to occupy an epistemic position that implies a Helmholtzian theory of perception. They borrow from Eliasmith the notion of a perspective, to be able to specify for each neuron or brain mechanism what information it has access to in a strict sense:A ‘perspective’, as I shall use the term, is a relation between an information processor and a transmitter of information. Perspective is determined by what information is available to an information processor from a transmitter (Eliasmith 2005, p. 99, italics in original).The notion of information implied here is just information in terms of (meaningless) energy transfer and should not be confused with intentional or semantic content. As Eliasmith puts it, the ‘perspective’ is a simple relation between a transmitter and a receiver of information. Assigning such perspectives is a useful tool for tracking information flows and covariance relations in complex systems like the brain. Eliasmith goes on to make the trivial sounding claim that “[A]n animal (and each of its information processing sub-components) can only access information available through sensory receptors” (ibid., p. 100). It is now a short step to the conclusion that since all the brain has direct access to are the changes that are taking place in its sensory registers, the only route to knowledge of what it is in the environment that is causing its changing sensory input must be through inference of some kind. Clark captures this type of reasoning in the following passage. We quote him in full since it is precisely this line of reasoning that we wish to challenge in this final part of our paper.

Clark writes:For, the task of the brain, when viewed from a certain distance, can seem impossible: it must discover information about the likely causes of impinging signals without any form of direct access to their source. Thus, consider a black box taking inputs from a complex external world. The box has input and output channels along which signals flow. But all that it “knows”, in any direct sense, are the ways its own states (e.g., spike trains) flow and alter. In that (restricted) sense, all the system has direct access to is its own states. The world itself is thus off-limits (though the box can, importantly, issue motor commands and await developments). The brain is one such black box. How, simply on the basis of patterns of changes in its own internal states, is it to alter and adapt its responses so as to tune itself to act as a useful node (one that merits its relatively huge metabolic expense) for the origination of adaptive responses? Notice how different this conception is to ones in which the problem is posed as one of establishing a mapping relation between environmental and inner states. The task is not to find such a mapping but to infer the nature of the signal source (the world) from just the varying input signal itself. (Clark 2013, p. 3)Clark seems to imply in this passage that there is a strong epistemic boundary separating the brain from the outside world. The brain has to infer the hidden state of the world that lies beyond the veil of sensory information. Any epistemic access to the world has to be inferential, because the brain has no direct access to the causes of its sensory information. Based on the flow and alteration of its spike trains, the brain needs to internally reconstruct the causal structure of the world outside.

This would seem to challenge the openness of the organism to the environment, suggesting instead that the organism is secluded and cut-off from the environment.29

We will argue that the boundary between the organism and the environment is real, but it is not the kind of boundary that hides causal structure in the environment from the organism.30 We make this argument from simulations of very minimal cases of life and cognition as discussed by Friston (2013a). In this paper, Friston shows two things. First, that coupled ergodic dynamical systems with a Markov blanket are a sufficient condition for free-energy minimization and second, that such systems will spontaneously occur in a simulation of a “primordial soup” with relatively arbitrary assumptions.31 The Markov assumption, as it is known from statistics, is a conditional independence of a node with its predecessors, given its immediate predecessors (or ‘parents’). For instance, in a domino chain reaction, the behavior of a given tile is only dependent on the behavior of its predecessor in the chain. Although the behavior of the predecessor is dependent on the rest of the causal chain, given the behavior of its predecessor, the domino of interest and the rest of the causal chain are statistically independent (Fig. 2).Fig. 2 Left A domino chain reaction as an exemplificiation of a Markov process. Every domino is only dependent on its previous domino, just as, in the figure on the right, every node is only dependent on its neighboring nodes. Right A schematic depiction of a Markov blanket, a spatial generalization of a Markov process. The gray circle represents the Markov blanket of a node, consisting of internal state (int), its children (the action states, act), its parents (sen), and parents of children (sen), with parent/child being understood in terms of cause/effect

The Markov blanket is an extension of the Markov assumption in the sense that it makes a node or a subgraph conditionally independent from the rest of the network given its neighboring nodes or ‘Markov blanket’ (Pearl 1988). Markov blankets are mainly used in the context of Bayesian graphical models, but as Friston (2013a) shows, they are all pervasive in natural systems dominated by short-range interactions. Due to the Markov blanket, Friston argues, a random dynamical system can be described naturally in terms of ‘internal’ and ‘external’ states ‘shielded off’ by a blanket of ‘perception’ and ‘action states’. This allows for an inferential understanding of coupled dynamical systems. Friston claims that, over time, the intrinsic dynamics of the internal states (something like electrochemical dynamics) will start to ‘infer’ the causally disconnected states of the intrinsic dynamics of the environment (that are ‘hidden’ behind the Markov blanket).

Markov blankets introduce what Eliasmith calls a perspective of an animal on its environment. Do Markov blankets imply the inferential seclusion of the animal from its environment? We think this does not follow at all, and rather points to a direct coupling between animal and environment. Friston’s simulation of the primordial soup shows that what he calls ‘inference’ is a natural consequence of processes governed by Newtonian and electrochemical interactions, completely understandable in terms of the basic and natural phenomenon of self-organization that Friston describes using the tools of graph theory. The notion of “inference” Friston is using is much more minimal than the one typically employed in philosophy, and refers to the mutual information (mutual predictability). Inference, for Friston, is the process that leads to nodes inside of a Markov blanket and nodes outside of the blanket having high mutual information. We understand inference as implying a probabilistic relation between representational states. However, in the primordial soup simulation, we doubt that it makes sense to ascribe representational states to the systems that are being modeled. Instead, we can explain the process that leads to mutual information more parimoniously in terms of synchronization.

The simulations in Friston (2013a) show that any dynamical system A coupled with another B can be said to “infer” the “hidden cause” of its “input” (the dynamics of B) when it reliably covaries with the dynamics of B and it is robust to the noise inherent in the coupling (Bruineberg and Rietveld 2014, p. 7). We question whether it makes sense to call this inference, given that the process that leads to high mutual information need not be thought of as representational. Friston recognizes that the internal and external dynamics of an animal in its environment are coupled through something called generalized synchrony. Generalized synchrony is a well-known phenomenon in physics, stemming from Huygens (1673) study on the synchronization of clocks. It is worth taking a closer look at this phenomenon very briefly because it is the minimal basis for his notion of inference as these concepts are understood by Friston in his writings on the free-energy principle.32 Fig. 3Original drawing from Huygens’ C. Horoloqium Oscilatorium. Two oscillating clocks hanging on a suspension beam that is itself able to move

Huygens observed that two pendulum clocks in the same housing start to synchronize through the tiny movement of the beam on which they are suspended, or the table on which they are placed. Such synchronization is a general and common phenomenon in non-linear dynamics (Strogatz 1994). When different time-scales are involved, one often speaks of ‘enslavement’, while when both systems exhibit similar typical times-scales one speaks of ‘entrainment’. One can understand these dynamics in terms of a synergetic ‘circular causality’ (Tschacher and Haken 2007). In systems with ‘circular causality’, there is no clear difference between internal dynamics attuning to external dynamics and vice versa (or clock 1 attuning to clock 2 and vice versa), rather both systems are bidirectionally coupled and reduce the disattunement between them until equilibrium (synchronization) is reached.

In this example the suspension beam, or table, acts as a Markov blanket. That is to say, the two clocks are statistically independent conditioned upon the movement of the beam. This is clear from the fact that no synchronization occurs when the beam is held fixed.33 When, for instance, the intrinsic dynamics of the two clocks are too different, no synchronization will occur.

Based on Friston (2013a), one would be required to give an inferential interpretation of the coupling of the two clocks, in which one clock “infers” the state of the other clock hidden behind the veil of the connecting beam. In this example, each clock is a ‘generative model’ of the dynamics of the environment (the other clock), coupled through the Markov blanket of the connecting beam. The synchronizing clocks are therefore excellent examples of being a model (discussed in the previous section). Clock A is a model of clock B if clock A resonates with clock B. Whether synchronization (or “inference”) occurs or not, is dependent on properties that have to do with the whole system: the period of the two clocks should not be too dissimilar; the beam should not be too rigid and not too flexible, and should be flexible in the right direction etc. We think such an inferential interpretation is unnecessary: there is no special inferential system inside of the Markov blanket, but the synchronization is the process of achieving high mutual information. This is emergent from the critically balanced coupled dynamics of the whole system.

Our point here is that, at least at this minimal level, there is no privileged explanatory role for the system that is defined ‘within’ the Markov blanket in achieving high mutual information with states of the environment. In the case of the two clocks, there is a perfect symmetry between ‘internal’ and ‘external’ dynamics. In the ‘primordial soup’ example of Friston (2013a), both the ‘organism’ and the ‘environment’ consist of essentially similar subsystems with electrochemical internal dynamics. It is only when these subsystems are similar enough that they will appear to exhibit active inference (i.e. to synchronize). Again, what counts as similar enough is dependent on the system as a whole.34 Given that attunement or synchronization takes place and given the ‘environment’, an observer can make predictions about the structure of the ‘agent’. In line with the free-energy principle, one can show that the average information-theoretic surprisal of the coupling is minimized when the dynamics of coupled systems are synchronized. In this sense, synchronization is a form of free-energy minimization, which can be given a Bayesian interpretation (the posterior divergence goes to 0 when the clocks synchronize). By synchronizing, clock 1 can be interpreted as maximising its Bayesian model evidence for the state of the environment. Although one can describe the behaviour of these systems in terms of probabilistic inference, this is unnecessary. The process of achieving high mutual information can better be understood in terms of the coupled dynamics of the system as a whole.