Building on the FEP, the postulation of Markov blankets, and the formal and conceptual description of variational active inference, the proponents of the extended active inference account, on which I will focus in this paper, suggest that it could “[…] provide a mechanistic basis for the psychological and functional aspects of the cognitive niche” (Constant et al. 2020, p. 9). Similarly, the closely related variational approach to niche construction, its proponents argue, “[…] could provide a promising modelling tool for research on the niche construction perspective” (Constant et al. 2018, p. 2). The epistemic goal of the extended active inference account and the variational approach to niche construction is, therefore, to capture the reciprocal causal relationship between an organism and its niche.
Referring to Stotz (2017) and West and King (1987), Constant et al. (2020) introduce a distinction between selective niche construction, developmental niche construction, and cognitive niche construction (equivalent to organism-niche coordination dynamics). However, they conflate this ontological and conceptual distinction: “The concept of the cognitive niche we refer to here is a sort of hybrid between the concepts of the selective, developmental, and cognitive niches” (p. 3). Unfortunately, they do not provide any reasons for this conflation. Similarly, in introducing their variational approach to niche construction, Constant et al. (2018) build on Stotz (2017) to establish a distinction between selective niche construction and developmental niche construction. However, they seem to misinterpret Stotz’s (2017) distinction by suggesting that “[the] developmental niche […] [is] a temporal subset of the selective niche” (Constant et al. 2018, p. 2), rather than a distinct causal pattern, that may or may not overlap with the causal pattern captured by selective niche construction. Furthermore, it remains unclear how exactly this distinction thus established translates into their variational model. Their expectation that “[…] the DNC–SNC [developmental niche construction–selective niche construction] tandem […] tend[s], over time and on average, to consolidate organism-niche complementarity across timescales” suggests that their model ultimately seems to conflate the ontologically and conceptually distinct processes of selective niche construction and developmental niche construction (Constant et al. 2018, p. 11). Given that selective niche construction, developmental niche construction, and organism-niche coordination dynamics capture different causal patterns with a different scope of the relevant dependence relations (see Sect. “Identifying causal patterns”), it is open to debate whether the extended active inference account and, to a lesser degree, the variational approach to niche construction can succeed in reaching their goal of describing the reciprocal relationship between organisms and their environment across scales. In what follows, I will focus on an analysis of the extended active inference account and enrich my observations and assumptions by taking the variational approach to niche construction into consideration.
Constant et al. (2020) are interested in the question as of how Variational Active Inference can contribute to a better understanding of “the computational function of cognitive extensions, and the (developmental and intergenerational) process whereby this computational function emerges” (p. 2). In exploring the relationship between an organism and its niche, they distinguish between the psychological habitat, defined as a space of action possibilities provided by the niche (i.e., affordances), and the functional habitat, defined as a space that provides epistemic resources for the completion of cognitive tasks. While the psychological habitat is set to explain the organism’s phenomenal experiences that are associated with its interaction with the niche, the explanatory role of the functional habitat is to elucidate the causal contributions of the niche to an organism’s ability to complete certain cognitive tasks. It is the functional habitat, understood as a sub-set of the overall physical habitat of organisms, that is key to understanding the functional role of the niche.
The reciprocal causal relationship between an organism and its niche, Constant et al. (2020) argue, can be aptly described in terms of variational active inference: “Changes in brain states and functional architectures optimize organisms’ generative (i.e., causal) model of the causal structure of their cognitive niche” in moments of perception (p. 9), thereby minimising free energy. Using the mathematical tools of variational calculus, this optimisation of the generative model through approximate Bayesian belief updating is formalised in the following way (Constant et al. 2020, p. 7):
$$ F\left( {s, \mu } \right) = D \left[ {q_{\mu } \left( \eta \right)\parallel p{(}\eta {|}s)} \right] - \ln p\left( s \right) $$
(1)
\(F\left(s, \mu \right)\) represents the free energy functional (the function of a function) of sensations \(s\) and internal states \(\mu \) of an organism, \(D\) represents the Kullback–Leibler divergence measuring the divergence between the variational density \({q}_{\mu }\left(\eta \right)\) and the distribution of the posterior probability of sensory causes \(p \left(\eta \right|s)\), which is always \(\ge 0\). The value of surprisal is expressed by \(-\mathrm{ln}p(s)\). This formalism captures the optimisation of an organism’s generative model in perception.
Conceptually, the optimisation of the model is modulated by the estimated precision of sensory signals. Precision estimation is understood as a “meta-learning mechanism” that facilitates the organism’s “attunement to causal regularities in the environment” (Constant et al. 2018, p. 9). Precision estimation is therefore crucial for optimising the bound on surprisal.
Importantly, while perception can optimise the bound on surprisal, only action can minimise the bound on surprisal given that it is “an attribute of sensations actively sampled” (Friston 2013a, p. 212). In active inference, proprioceptive and interoceptive changes “[…] resolve uncertainty through the active sampling of salient, uncertainty reducing sensations, while allowing for preferred, unsurprising outcomes” (Constant et al. 2020, p. 8; see also Hesp et al. 2019). This is captured by the following formal expression (ibid.):
$$F\left(s, \mu \right)=D\left[{q}_{\mu }\left(\eta \right)\parallel p(\eta )\right]-{E}_{q}\left[\mathrm{ln}p\left(s\left(a\right)|\eta \right)\right]$$
(2)
Action, this expression suggests, minimises surprisal by sampling the niche in such a way that it conforms to expected sensory and internal states. Active sampling is guided by the salience of specific causal regularities in the niche. The salience of certain objects and artefacts, for example, structures the active minimisation of free energy (Constant et al. 2018). Salience, then, is a relational property establishing connections between the organismic generative model and components of the generative process, i.e., the niche. The salience of causal regularities in the niche is, in many cases, a result of previous niche construction activities. Constant et al. (2018, 2020) describe the process of actively creating salient causal regularities so as to facilitate free energy minimisation as uploading.
In the context of analysing cognitive cases of free energy minimisation, Constant et al. (2020) suggest that the notion of cognitive uploading is an extension of the concept of cognitive offloading, where cognitive offloading is defined as “the use of physical action to alter the information requirements of a task so as to reduce cognitive demand” (Risko and Gilbert 2016, p. 676). Offloading refers to the manipulation of objects in the niche by an individual organism so as to complete a specific cognitive task (e.g., remembering, problem-solving). In cases of uploading, Constant et al. (2020) propose, “[a] function is uploaded when social and technological change means it is now taken care of by the niche rather than the individual” (p. 5). I will return to the relevance of uploading for understanding variational active inference in the Sect. “Extended active inference and organism-niche coordination dynamics”. For now, it suffices to note that salient causal regularities are, at least in many cases, the result of cognitive uploading. In turn, as Constant et al. (2020) argue, “[…] cognitive uploading helps agents to minimize the bound on surprisal” (p. 15).
So far, I have summarised Constant et al. (2020) conceptual and formal description of the ways in which an organism minimises its free energy by generatively modelling its niche (i.e., the generative process) in perception and action. As seen in previous section, free energy minimisation is dependent upon the statistical separation between internal and external states by a Markov blanket. The Markov blanket induces a formal symmetry between the internal states of the organism and the external states of the niche. Given this formal symmetry, it is possible to transform the variational formalism such that the niche becomes the generative model and the organism becomes the generative process (Constant et al. 2018, 2020). In other words, “[…] there must be a description of niche dynamics, where the environment models how the agent’s internal dynamics are generating its active states” (Constant et al. 2018, p. 5). The consequence of this transformation is that the organism’s active states are represented as the sensory states of the niche, while the sensory states of the organism become the active states of the niche (cf. ibid.). This circular causality (Bruineberg et al. 2018) of the states of the organism and its niche can be expressed by rewriting Eq. (1) so that the sensations \(s\) and internal states \(\mu \) of the organism in Eq. (1) become the organism’s actions \(a\) as ‘sensed’ by the niche and the states of the niche \(\eta \) (Constant et al. 2020, p. 10):
$$F\left(a, \eta \right)=D \left[{q}_{\eta }\left(\mu \right)\parallel p\left(\mu \right|a)\right]-\mathrm{ln}p(a)$$
(3)
This formal transformation enables a description of the causal relationship between an organism and its niche ‘from the perspective’ of the niche. More importantly, however, it allows us to describe the organism and its niche as a causally coupled system that jointly minimises free energy. As Constant et al. (2018) put it, “[…] the generative model of both agent and niche are the veridical explanations of their respective sensory fluctuations and the variational free energies of both are jointly minimised with respect to the posterior beliefs parametrized by their respective states” (p. 7).
The assumption that the reciprocal causal relationship between an organism and its niche can be described in terms of joint free energy minimisation finds support in a recent simulation study of desire path behaviour, where this behaviour is a simplified example of niche construction (Bruineberg et al. 2018). In a nutshell, these simulations show how an organism and its niche (a simple maze) are mutually attuned to each other through the organism’s actions (the carving of a path through the maze), which feed back into the organism’s sensations of the niche, thereby jointly minimising free energy.
On this view, free energy minimisation is conceptualised as a process that comprises the coupled organism-niche system. In this sense, Constant et al. (2020) argue, free energy minimisation is extended: “[…] the free energy of two systems (i.e., organism and niche) is just the sum of their respective free energies, conditioned upon the (i.e., sensory and active) states they share (Bruineberg et al. 2018)” (p. 12). Under this description, active inference extends into the niche. The upshot of Constant et al. (2020) account is a theoretical integration of Variational Active Inference and philosophical work on extended cognition (Clark 2008; Clark and Chalmers 1998; Kirchhoff 2012, 2015). I will discuss the feasibility and implications of this integration in detail in the next section.
For now, the question arises how the extended active inference account relates to the distinction, introduced above, between selective niche construction, developmental niche construction, and organism-niche coordination dynamics. These three concepts, I have argued above, capture different causal patterns with a distinct scope. Recall that Constant et al. (2020) state that “[t]he concept of the cognitive niche we refer to here is a sort of hybrid between the concepts of the selective, developmental, and cognitive niches” (p. 2). This conceptual hybridisation, however, conflates important ontological and conceptual differences between selective niche construction, developmental niche construction, and organism-niche coordination dynamics. This conflation would render the variational model of the reciprocal causal relationship between an organism and its niche unconvincing, as this model would not capture the different causal patterns – and their inter-relations – characterising the target phenomenon. This conclusion, however, could be avoided if we re-assess the niche conception that is at the core of the extended active inference account. It is possible, I propose, to appreciate the explanatory merits of the extended active inference account without committing to the conceptual conflation of selective niche construction, developmental niche construction, and organism-niche coordination dynamics.
To explore this possibility, I will take a closer look at the variational model suggested by Constant et al. (2020) and interpret it against the background of the distinction between the different causal patterns that contribute to the target phenomenon, i.e., to the reciprocal relationship between organisms and their environment. The formal description of extended active inference, I argue, represents the causal reciprocal relationship between an individual organism and its cognitive niche. Equations (1) and (2) above have it that it is the individual sensing and acting organism that generatively models the niche. Given the formal symmetry between the internal states of the organism and the external states of the niche induced by the Markov blanket, the states that realise extended active inference, captured in Eq. (3), are, likewise, the states of an individual organism and its cognitive niche. This joint minimisation of free energy, my interpretation of the formal expression of extended active inference implies, is realised in the here-and-now or at times t1, t2, t3, …, if we wish to describe the temporal unfolding of extended active inference. As it stands, the extended active inference account does not capture selective niche construction, for it remains silent on the selection pressures induced by niche construction activities by an entire population across multiple generations. Furthermore, the extended active inference account does not capture developmental niche construction, because its focus is on the individual organism, rather than on an entire population (or sub-population) of organisms that modifies the cognitive niche across two (or more) generations, thereby changing the array of possible phenotypic traits. Contrary to the stipulation by Constant et al. (2020), then, extended active inference does not capture “a sort of hybrid between the concepts of the selective, developmental, and cognitive niches” (p. 2), but organism-niche coordination dynamics.
This conclusion can be further substantiated by taking the assumptions of the original extended cognition thesis into consideration. Recall that the analysis of cognitive niche construction developed by Clark (2008) focuses on the engagement of an individual organism with resources in the cognitive niche (see Sterelny 2010). Furthermore, in their original proposal of the extended cognition thesis, Clark and Chalmers (1998) stress that the cognitive resources in the niche “[…] are active, playing a crucial role in the here-and-now” (p. 9; italics in original). This way of theorising identifies the individual organism in its occurrent interaction with the cognitive niche as the relevant causal pattern for explaining cognition. Given that the extended active inference account is committed to these key principles of extended cognition, it follows that this account captures organism-niche coordination dynamics, rather than selective niche construction or developmental niche construction.
A similar interpretation of the explanatory scope of Variational Active Inference applies, I suggest, to Constant’s et al. (2018) variational approach to niche construction.Footnote 4 Their formal model of the reciprocal causal organism-niche relationship, and joint free energy minimisation, bears a strong resemblance to the model developed by the extended active inference account. Their model, it seems to me, captures organism-environment coordination dynamics, rather than selective niche construction or developmental niche construction. This becomes obvious in their analysis of situated learning under variational active inference, for example, which provides a formal and conceptual interpretation of the ways in which an organismic generative model is updated so as to be optimally attuned to the generative process, i.e., the niche. According to Constant et al. (2018), situated learning
[…] requires of organisms that they engage conventionalized patterns of collective activity, grounded in material artefacts and other physical aspects of the niche, which have been transmitted to allow for the progressive integration of the organism as a legitimate member of the community. […] By providing opportunities for learning how to learn, the developmental niche does not merely indicate what information ought to be learned, but also how best to learn it. As a meta-learning device, the information encoded in the material states of the local environment function to guide active inference by weighting sensory inputs according to their reliability, or salience, and constitute a more general device that allow organisms to learn how to learn, by guiding action–perception cycles (active inference). (p. 9)
Clearly, active inference, precision estimation, and the active exploitation of salient causal regularities in the niche, as they are described by Constant et al. (2018), are features of organism-niche coordination dynamics. The reason for this is that “action–perception cycles”, precision estimation, and salience exploitation are realised by an individual organism, rather than by a population of organisms, in the here-and-now and across time points t1, t2, t3, …, not across the lifetimes of multiple generations (in the case of selective niche construction) or across the lifetimes of at least two generations (in the case of developmental niche construction), as I have argued above. The developmental niche certainly plays a role in organism-niche coordination dynamics, and this role is captured by Constant’s et al. (2018) treatment. However, what is not captured by this approach is the population-specific, trans-generational, cumulative process of developmental niche construction, even though Constant et al. (2018) might claim otherwise. Furthermore, the variational model suggested by Constant et al. (2018) does not represent the causal pattern of selective niche construction.
My claim that the variational model, as it stands, exclusively captures organism-niche coordination dynamics, and leaves the causal patterns of selective niche construction and developmental niche construction unexplained, can be further substantiated by noting how Constant et al. (2018) summarise their approach:
Active inference allows the organism to specify (often implicitly) those features of the environment that will be adaptive given the demands of their phenotype, i.e. they can specify those features of the niche, the learning of which makes them an accurate model of their environment. On a developmental timescale, this rests on the meta-learning function of niche construction, which guides optimization, or phenotypic plasticity. On the scale of phylogeny, it rests on the inheritance of constraints passed on across generations in the form of species relevant information (salience), which will further guide the optimization in development. (p. 11; emphasis added)
I take it as uncontroversial that organism-niche coordination dynamics, which is captured by the variational model, “rests on” population-specific, cross-generational processes of selective niche construction and developmental niche construction. Yet, I do maintain that organism-niche coordination dynamics, selective niche construction, and developmental niche construction are characterised by different causal patterns, each with a different scope.
The upshot is that the extended active inference account (Constant et al. 2020) and the variational approach to niche construction (Constant et al. 2018) provide variational models of organism-niche coordination dynamics. However, these models do not represent the causal patterns at play in selective niche construction and developmental niche construction.
In reply to this conclusion, the proponents of the extended active inference account (Constant et al. 2020) and the variational approach to niche construction (Constant et al. 2018) have three options. First, they could deny that there is a non-trivial ontological and conceptual distinction between selective niche construction, developmental niche construction, and organism-niche coordination dynamics. They would then have to show why this distinction is not relevant for their variational models and why and how the extended active inference account and the variational approach to niche construction do capture all three causal patterns.
Second, they could (at least to a certain extent) agree with my analysis and suggest that the reliance of their models on the Markov blanket formalism allows them to extend their model so as to capture selective niche construction and developmental niche construction as I have characterised it. According to variational neuroethology (Ramstead et al. 2019), a recent framework based on Variational Active Inference, the general definition and formalisation of Markov blankets allows for the possibility of recursively nested Markov blankets. The idea is that “[…] the Markov blanket ontology can be reiterated recursively, such that the MBs [Markov blankets] at any one scale are composed in turn of MBs at the scale above and below – which are also made of MBs, and so on, all the way up and all the way down” (Ramstead et al. 2019, p. 192; see also Hesp et al. 2019). They could then attempt to show, on formal and conceptual levels, that selective niche construction and developmental niche construction could be captured by modelling them as processes defined over different ensembles of recursively nested Markov blankets. This strategy, if the resulting formalisations were mathematically and computationally tractable, could save the general idea that Variational Active Inference has the resources to capture selective niche construction, developmental niche construction, and organism-niche coordination dynamics while avoiding the limitations of the current versions of the extended active inference account and the variational approach to niche construction.
Finally, a third option for replying to the conclusion of my analysis would be to agree that the extended active inference account and the variational approach to niche construction only capture organism-niche coordination dynamics and allow for the possibility that different explanatory styles might be needed for understanding selective niche construction and developmental niche construction, for example non-variational mathematical models in evolutionary biology (Laland et al. 2000) and developmental systems theory (Griffiths and Hochman 2015; Griffiths and Stotz 2018), respectively. This would entail a commitment to “pluralism of explanatory styles” (Potochnik and Sanches de Oliveira 2020, p. 1314).
These three options offer a standing invitation to the proponents of the extended active inference account and the variational approach to niche construction to reply to my analysis. For the time being, however, I submit that the extended active inference account, against the key assumption made by its proponents, merely captures organism-niche coordination dynamics and leaves the question as of how selective niche construction and developmental niche construction can be explained in terms of variational active inference unresolved. Still, this allows for the possibility to examine the integration of Variational Active Inference with philosophical theorising about extended cognition. I will take up this task in the next section.