A variational approach to niche construction

Youngsters spend a great deal of time watching and imitating adults' nut cracking prowess with smaller stones and nut pieces [16]. Adults, in turn, leave material behind them, and, importantly, different sorts of reliable ecological traces like oily residues on the stone anvil site, and pieces of shell, as well as the hammer stones used for nut cracking, which further attract and direct youngsters’ interest to the specifics of the nut cracking site.

These alterations to the environment last for years, even when exposed to inclement climates. Although the learning of the nut cracking technique is partly mediated by attending to adults and imitating them, this, in itself, is not sufficient for the learning. Rather, it is the high fidelity, lasting changes to the niche and the technical artefacts left behind by adults that indirectly biases learning over short and long timescales, thus supporting persistent practice over the course of development [16].

Lasting changes made to the developmental niche enable the acquisition of complex nut cracking skills in youngster Capuchin monkeys, and supports the passing on of traditions, defined as the renewed learning of behavioural patterns by each new generation [36,87]. Crucially, this depends upon the changes to the niche which extend the reach of knowledge transmission through the inheritance of salient information over intergenerational timescales.

In turn, such information transmission depends on two interrelated processes associated to DNC and SNC: (i) the construction, engagement, and re-engagement of the developmental niche, along with the information it comprises; (ii) the slow learning fast inference differential between the environment, and the organism (cf. §4.2.). Engaging the niche leads to the accumulation, e.g. of oily residues and nutshells, which increases the precision of the environment understood as the propensity to influence learning. A precise environment is likely to be passed on to the next generation, and thereby can extend the reach of information transmission.

The sort of (non-genetic) niche inheritance that we describe here is a form of ecological inheritance, which rests on the interplay of two kinds of resources: (i) algorithmic information, and (ii) material and energetic resources [88]. Algorithmic information consists of ‘know-how’-like information encoded by structural and functional features of the niche (typically, genetic information or direct knowledge transmission) [25], which is passed on through standard niche inheritance. Material and energetic resources typically correspond to abiota (e.g. material resources like oily residue). Taken in the context of evolutionary biology (and ecology cf. [89]), algorithmic information is defined as anything that can reduce organisms' uncertainty with regard to possible fitness advantage afforded by the selective environment [90,91]. It eases the exploitation of the environment, and in so doing, can causally influence reproductive success [91]. Crucially, on that view, abiota passed on through ecological inheritance can be a source of algorithmic information, e.g. knowing how to exploit food resources based on the traces left by the activity of members of—current and—previous generations. Following the variational free-energy perspective, we can further define algorithmic information as salient information that secures the optimization (e.g. learning and phenotypic plasticity) of the phenotype throughout development; a meta-learning device that members of previous generations craft through reiterated active inference.

The challenge with algorithmic information is that while its transmission is costly, it is nonetheless a necessary resource [91]. The reproduction of the adaptive life cycle that is secured through development involves myriad environmental manipulations (e.g. building nests, beaver dams and knocking palm nuts), which often require organisms to know-how to engage their environment early in development. Organisms can be informed a priori, for instance, through algorithmic information that is bound up with their genetic makeup, but also by having access to external channels of algorithmic information. The inheritance of algorithmic information supposes that previous generations themselves had access to algorithmic information having enabled them, in the first place, to exploit the resources of their environment, so as to be efficient enough to afford passing on this information. The inheritance of algorithmic information is a costly business, as it necessitates of previous generation that they invest metabolic resources for its physical acquisition, storage and use, and transmission [91].

From the point of view of the variational approach, however, algorithmic information, understood as salient information, comes at low cost, as it carries the cognitive burden of precision estimation in development (cf. end of §4.3.), which is one of the main mechanism underwriting organisms’ homeostasis (cf. §3.3). Therefore, because the material and energetic carriers of algorithmic information enable cheaper sensory information processing, algorithmic information can compensate for its cost of transmission. As salient information, it could be modelled as environmental precision, and its role in learning could be explored in simulations similar to the one presented in §4.2. It can be approached quantitatively as information gain cf. §3.3., 4.1.). One could, for instance, test the function of algorithmic information by comparing the changes in variational free energy over time between different organism–niche joint phenotypic spaces, in which agents' differential saliences could vary within and across trials. Variations within trials would represent learning in development (DNC scale), and variation across trials would represent the inheritance of more or less precise algorithmic information (SNC scale).

We conclude our discussion by considering the challenges to NCT presented earlier in light of the variational approach to niche construction (see figure 4 for a visual summary). We examine some of the predictions of the variational approach to niche construction with regard to (i) the generation of non-random, organism-dependent biases on selection pressures; (ii) the consolidation of organism–niche complementarity across temporal scales; and (iii) the cross-generational niche stabilization of environmental conditions.

(i) As a corollary of active inference, niche construction has a systematic causal influence on fitness, on average and over time, as it optimizes the attunement of the organism with its local environment. This is especially important in organisms for which fitness depends on the ability to navigate and attune to complex social environments, and participate in immersive patterned cultural practices, such as humans [37]. Niche construction induces changes to the environment that are not (necessarily) random. On the contrary, many of these changes (although they may be implicit) are continuous with the organism's expectations (the set of priors embodied by the organism), as they result from active inference, which is controlled by these expectations. We can further predict that some of these changes will be deeply continuous with the environmental selection pressures themselves, as these are ‘expected’ by the phenotype of the organism. This is so because natural selection retains sets of priors best adapted to pressures in first place ([20–24] also see electronic supplementary material), which active inference brings forth.

(ii) 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. Therefore, we can expect the DNC–SNC tandem to tend, over time and on average, to consolidate organism–niche complementarity across timescales.

(iii) The optimization of organism–niche complementarity entails the cross-generational stabilization of the features of the niche, which can be framed as the consolidation of salience: the more agents implicitly engage a material locus of information, the more likely this locus is to further attract engagement, and consequently, the more salience it can acquire. Because of the nature of active inference, organisms tend to learn from (adapt their model to) highly salient environments, potentially deploying culturally patterned practices built around these. Interestingly, we can speculate that an environment that acquires too much salience will become maladaptive, as the attunement dynamics taking place in development require that both the organism and the environment share a certain propensity to learn from one another (c.f. figure 4).

The FEP has motivated the production of much research over the last 10 years. While this work has led to novel theoretical developments, ecologically valid empirical work is forthcoming still. The approach we propose in this paper faces the same limits. So far, it can only be used to make predictions, one of these being that free energy can be cast as an ecological quantity reporting organism–niche complementarity. Much work is still to be done to demonstrate the validity of the variational approach to niche construction. Nonetheless, simulation studies can be used to explore the ecological phase space and to generate predictions about adaptive solutions to ecological problems, framed as free-energy bounding problems. Scientists can proceed on the basis of such predictions and compare these computational models with empirical data. We can further draw from this approach an important conceptual point, which can inform theorizing and empirical work: over time, part of the architecture of priors embodied by members of a population (brain- and body-based priors) becomes encoded in the socio-material setting of the niche. The organism then models an environment which, in turn, models it back, and thereby living systems end up expecting a world that reflects their own expectations. This, however, points to another limit of the variational approach. Thus far, it has only focused on positive feedbacks involved in niche construction, in relation to the adaptation of individual species. Future research should consider the integration of negative feedbacks caused by niche construction, and how those play out in the dynamical relation between natural selection and niche construction, and their consequence in eco-evolution. Moreover, future research should consider the possible limitations of the variational approach with regard to the complexity of ecological inheritance.

For instance, modelling the full scope of ecological inheritance might become quickly intractable due to the complexity of the formalism of the variational approach.

We thank William-J. Beauchemin, Jelle Bruineberg, Laurence Kirmayer, Jonathan St-Onge and Sander Van de Cruys for helpful discussions and comments.

A.C. conceived of the architecture of the text and the main arguments (over all the sections). He carried out most part of the data interpretation in relation to the concepts discussed in the paper and drafted most part of the manuscript. M.R. carried an important part of the conceptualization work in the paper and helped draft parts of the manuscript, especially that which relate to the discussion of the Variational Free Energy Principle (§2). S.V. carried an important part of the conceptualization work in the paper, especially that which relate to (§4). J.C. was added to help us with the fundamental concepts in biology, at the request of a member of the revision board. He also helped draft throughout the text. He also wrote the supplementary material. K.F. performed the simulation and helped draft the manuscript, especially with regard to the mathematical formalism (§3). Importantly, all authors participated substantially in the two rounds of revision. All authors gave final approval for publication."

Work on this review was supported by the Social Sciences and Humanities Research Council of Canada (M.J.D.R.) and the Wellcome Trust, Principal Research Fellowship (K. Friston) ref. no. 088130/Z/09/Z.