From cognitivism to autopoiesis: towards a computational framework for the embodied mind

It is this ‘just so’ circularity that the FEP seeks to resolve by appeal to enactivism. Rather than the post-hoc definition of priors or cost functions, the FEP derives a normative, a priori first principle from a provable definition of living systems. To do so, the FEP highlights the necessary tendency of living organisms to resist the second law of thermodynamics; i.e., to maintain an internal structure or dynamics in the face of constant change. That is to say, by definition, living beings are those that maintain an upper bound on the entropy of their possible states. One can see this by considering a candle flame or snowflake; although both have some degree of persistent local dynamics, these do not resist the constant flux of the physical universe; they instead dissipate rapidly in the face of environmental fluctuations (a gust of air or the warmth of the sun). In contrast, to live is to visit some states more frequently than others—and visit their neighbourhoods time and time again (for example, our daily routine). However, before these imperatives can even be considered, the very existence of a system mandates the separation between the system and its external milieu (e.g., the environment in an evolutionary setting or heat bath in statistical physics). It is the separation or boundary that lies at the heart of the enactivist imperatives for predictive processing.

For example, a cell persists in virtue of its ability to create and maintain a boundary (cell-surface), through which it interacts with the environment, thereby maintaining the integrity of the boundary. It is this autopoiesis, or self-creation, which enables the system to limit the possible states it visits, and thus to survive (Varela et al. 1974). The FEP recasts this as a kind of self-fulfilling prophecy, in which an organism itself constitutes, in the generative sense, a belief that it will prevail within certain embodied and environmental conditions. In short, the very existence of a system depends upon conserving its boundary, known technically as a Markov blanket, so that it remains distinguishable from its environment—into which it would otherwise dissipate. The computational ‘function’ of the organism is here fundamentally and inescapably bound up into the kind of living being the organism is, and the kinds of neighbourhoods it must inhabit.

From this fundamental property of existence, it follows that any biological organism will possess the following characteristics: Ergodicity By placing an upper bound on entropy, an organism will necessarily occupy (the neighbourhood of) some states more often than others. This means that the average probability of being in a given state is equal to the probability of the system being in that state when observed at random. Note that this is simply a reformulation of the overall principle; to live (resp. be) is to revisit (resp. occupy) some characteristic states over time. A Markov blanket The boundary (e.g., between internal and external states of the system) can be described as a Markov blanket. The blanket separates external (hidden) from the internal states of an organism, where the blanket per se can be divided into sensory (caused by external) and active (caused by internal) states. Active inference The Markov blanket induces a circular causality because sensory states depend on hidden states that depend on active states, which depend upon internal states. In other words, the sensory and active states (that constitute the Markov blanket) mediate perception and action that are locked into a perpetual cycle to upper bound the entropy of both. Because the entropy of the Markov blanket is, by ergodicity, the time average of surprise or negative Bayesian log model evidence, sensory and active states will appear to maximise Bayesian model evidence. This means internal states can always be cast as representing external (hidden) causes—and thereby constitute a generative model of the causal forces that impinge upon them—while active states change the external states to make this job easier (e.g., avoid dark rooms).

Autopoiesis Because active states change external (hidden) states, but are not changed by them, they will place an upper (free energy) bound on the entropy of biological states. This is because they are caused by internal states, and will therefore appear to maintain the structural and functional integrity of the internal states and their Markov blanket.Simply put, an organism persists in virtue of having internal states which cause surprise-minimizing, evidence maximising actions; these in turn maintain the partitions described above, which is a necessary precondition for existence: c.f., the self-evidencing brain (Hohwy 2016). One can formulate this in another way; the organism’s internal states constitute probabilistic beliefs about what actions are the most likely to provide evidence for the organism’s existence (survival). My actions are not merely the output of an internal dynamic; the FEP argues that if I am to survive, they will actively bring about the conditions for my survival. The point is that the boundary itself is constituted by an ergodic dynamical interchange between ‘internal’ and ‘external’, rather than a cognitivist predominance of internal processing.

This notion is at the heart of autopoietic views of life and mind, insofar as it induces a deeply circular causality between internal and external states, to provide a normative principle by which to understand all action and perception. If an organism is endowed with the belief that it will maximize the evidence for its existence, then it will act in ways that are consistent with that belief. In other words, if survival is synonymous with minimizing surprise—i.e. maximizing evidence or self-evidencing (Hohwy 2016)—then it follows that the only possible prior belief an agent can entertain is that it will behave so as to minimize surprise. This is easy to see through reductio ad absurdum: if I believe I will be surprised, the only way I can be surprised is if I am not surprised. More exactly, the organism, body-brain-and-world itself constitutes the ‘belief’ or generative model that it will survive; in a very concrete sense, the kinds of limbs and morphological shape one has will constrain the probabilities of the kinds of actions one can engage in. This can be considered by analogy to the notion of an Umwelt, in which an organism’s world is itself a constituting and constraining feature of its embodiment (e.g., the isomorphism between the wavelength selectivity of our photoreceptors and ambient radiation from the sun).

This deep reciprocity between the embodied and environmental facts of the organism is embedded in the pattern of neural patterns which preconfigure the entity to best survive within its living world. Even seemingly ‘representationally hungry’ operations will be enmeshed within these looping, self-sustaining dynamics. For example, the organism will choose options that minimize its surprise, where free energy provides a tractable bound on surprise; hence the FEP. This bound is not absolute, computed solely in the head, but instead relative to the embodied nature of the organism as selected (via evolution) by the type of body and environmental niche inhabited by the organism. The implication is that my internal representations—the generative model of the world embodied in the web of neural connections—are causally coupled to my homeostatic needs and the environmental niche within which my brain has evolved.

Heuristically, this means that I will behave in ways consistent with my survival—which is itself consistent with or constrained by the type of body that I have, the econiche within which I have co-evolved, etc. If I am a cave bat, I will hang around in dark caves. If I am a human being, I will seek out other human beings and read articles on philosophy. The body itself is thus a prior boundary condition, or a conditioning factor, in the overall generative model defined by my Markov blanket. My body directly shapes my possibilities for (active) inference. The body-brain system has evolved to constitute a generative model, which specifies the types of behaviours, and environments in which I am likely to engage. Where one draws the boundaries is a matter of the question one wishes to ask; any living organism will be defined by a nearly infinite matryoshka embedding of blankets-within-blankets.21

The FEP thus provides a formal, information theoretic framework within which to explain the constitutive coupling of the brain to the body and the environment. The ‘cost function’ or imperative priors arise directly from the interoceptive, homeostatic needs of the body in exchange with the environment. My brain and body themselves constitutes a ‘belief’, in the generative sense about the kinds of states (e.g., homeostatic set points such as temperature, blood glucose) I must inhabit if I am to survive.22 The imperative to reduce free energy renders any action, which improves my survival inherently ‘desirable’—in the sense it brings me back to the attracting states prescribed by my generative model. Where, crucially, my self-evidencing generative model is learned or inherited from the environment; the capacities of my limbs for action preconfigure the nature of my active-inference.

Clearly, the active inference account satisfies the criteria for a radically embodied theory of mind. According to the free energy principle, an organism is best understood as a system of mutually interlocking systems; the body, mind and environment are inextricably bound up in the organism’s free energy minimization: in fact, all the heavy lifting done by active inference is in preserving a degree of (statistical) separation between the body, mind and environment (by maintaining the integrity of their respective Markov blankets). Perception is enactive and affective; through the reduction of surprise and uncertainty, perceptual and active states are selected to maximize the evidence for my existence. The body itself is a part and parcel of the computational machinery that leads to my survival. By elucidating these principles down in a formal, computational framework, the FEP provides an understanding of these issues that is amenable to experimentation and formal analysis. Although the FEP provides a normative, teleological essence to the synthesis of biology and information, the specifics of compliant (neuronal and behavioural) process theories must be discovered and verified empirically.

This is because, as a process theory, the specific couplings of action and body are left unspecified; which systems in the brain encode the uncertainty of some cognitive domain? What are the functional dissociations themselves; e.g., what solution has nature found to optimize the brain-body-econiche ensemble? In what specific ways do the affordances disclosed by these relations impact the cortical hierarchy, and vice-versa (Bruineberg and Rietveld 2014)? By providing much needed guideline to discovery (Chemero 2011), FEP renders a productive union of the embodied cognition and information theory, allowing the enactivist not only to describe the importance of the body, but to also build models of the brain-body-world relationship (See Friston et al. 2012a, b for one illustrative example).

FEP or active inference does not do this job for free; rather it provides a state theory, under which to develop specific process theories. One might here ask; if the FEP is unfalsifiable—in the sense that Hamilton’s principle of least action is not, in itself, falsifiable—is it uninformative? The FEP is uninformative in the sense that the principles of natural selection do not explain a particular species or phenotype; however, they inform the viability and sufficiency of any process theories. For example, expected utility theory and reinforcement learning are not sufficient theories of behaviour because they do not link utility or reward to free energy. Conversely, the theory of evolution by natural selection is free energy compliant (through its formulation as Bayesian model selection).

We note however, that not all are likely to agree with this interpretation. Some theorists take the Markov blanket to imply a strong partitioning, and this has led arguments directly contravening this type of claim; if my survival depends upon separating internal from external states, then I must infer through mental representation (i.e., Bayesian inference) the causal nature of those hidden, external states (Hohwy 2016).

There are at least two responses to this; first, the aim of the FEP, and the account sketched here, is not to deny the causal importance of the internal states. Indeed, an organism survives in virtue of its nervous system constituting (a partition of) a generative model, which can infer those actions most likely to maximize the evidence for its survival, as sampled indirectly through sensations. However, the delineation of the boundaries of a Markov blanket is essentially relative and variable; indeed any organism will be defined not by a singular blanket, but instead by a near-infinite regress of causally interacting Markov blankets within Markov blankets (and indeed of Markov blankets). In other words, the brain does not constitute a single blanket, but rather innumerable systems that are at once modular and dynamic, creating internal states in virtue of their interactions with one another. There will be a brain-body blanket, which describes the same circular causality. One can go still further; in an ecological sense, my ‘econiche’ or self-creating environment of tools and cultural settings—and my interaction within it—constitutes another blanket (See Bruineberg et al., this issue). Furthermore, blankets of blankets may change over many timescales; for example, with developmental cycles. This begs interesting questions about the timescales over which ergodicity applies—and whether these timescales are nested. Where one draws the causal demarcation is simply a matter of the question at hand. At all levels of description there will be a constant interaction between an emergent generative model and a coupled interaction thereof.

In short, the Markov blanket does not provide a cover from which to hide from external states in a radically sceptical fashion. On the other hand, the Markov blanket does not admit a radically realist position; in the sense a living system can never know what is ‘out there’—it can only infer, with a greater or lesser degree of accuracy, the causes of sensory impressions on the blanket. Perhaps it is best to construe Markov blankets as ‘gluing’ the brain to its body and the body to its econiche; however, by definition, this glue never comes with ‘phenomenologically transparent’ on the tin. This is not to say that the Markov blanket precludes direct perception in the sense of delivering an ‘optimal grip’ on the affordances offered by inferred external (hidden) states (Bruineberg and Rietveld 2014). In active inference, everything is in the service of predicting action, no matter how deep the hierarchical processing.23

Clearly, the explanatory scope of the FEP goes far beyond that of B-representations. Bodily information, and inference over those signals, is not only a contextual contributor. Rather, it is a part of the causal tapestry, which defines the agent and its viability as a ‘model’ of its environment. Another way to say this is: the causal machinery of the brain and its representations are enslaved within the brain-body-environment loop of autopoiesis—which is reminiscent of the circular causality that underwrites the slaving principle in synergetics (Haken 1977). This is an important point here; the FEP is not eliminating representations,24 nor is it hiding the body within the brain. Rather, the FEP directly explains the organism: the embodied brain becomes a model of its environment in virtue of exactly the kinds of autopoietic, dynamics espoused by proponents of radical embodiment.

Predictive processing promises to revolutionize our understanding of how the brain—in conjunction with the body and environment—persists in the face of uncertainty. Here, we have reviewed a continuum of views; all of which can be gathered under predictive processing. Our intention was not to exhaustively cover all possible contributions; given the exponential growth of prediction-related theoretical and empirical work, such an effort would beyond the scope of the present issue. Instead, we have focused on how—rather than constituting a single, monolithic framework—predictive processing involves a wide range of theoretical commitments, with respect to the philosophy of mind. As we have shown, depending upon where one falls along this spectrum, competing modular, connectionist, or embodied/enactive theories can easily arise. Nevertheless, the FEP accommodates many of these disparate views within a single framework, whilst motivating novel empirical work.

Here we have sketched the enactive and embodied underpinnings of the FEP. Much work remains to be done on this issue. For example, although we here argue that the FEP resolves the tension between internalist and externalist approaches, we recognize that some may remain dissatisfied. We look forward to discourse concerning, for example, the notion of the brain-body-econiche as a kind of generative model, and whether one can sensibly speak of ‘representations’ under such a framework. If representation is distributed all around the brain and beyond, does it remain a representation in any meaningful sense? Similarly, many aspects of predictive processing suggest a role for ‘emergent’ or global processes (See Kim 1999 for critique), which constraints local processes in a kind of hermeneutic circle. It remains to be seen whether these are at tension with the account given here, and whether they lend themselves to a more cognitivist or representational view. Again, rather than delve deeply into these issues, our intention as to instead paint a general picture. For those who are interested, this special issue contains substantive views arguing for example, that the FEP does not require any ‘inference’ or ‘internal models’ (Bruineberg et al. this issue).

In summary, the FEP offers a formal path forward for enactivism. By providing a guideline to discovery, the normative principles embedded within the approach allow enactivists to go beyond arguing about the demarcations of the organism, to instead develop empirical theories of how brain, body, and world interact with one another. Ultimately it is this pragmatic ability to motivate testable hypothesis about ‘enactive computation’ that may most benefit cognitive science.