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

One might object that there is still a non-trivial boundary separating the system from its environment in both the examples discussed above. Both systems are causally interacting through a Markov blanket. We agree, but we do not think this has any implications that conflict with the arguments of this paper. The importance of such a boundary for living organisms has been central in the autopoietic approach from the very start (Maturana and Varela 1980; Varela et al. 1991). The mathematical methods used in ecological psychology also rely heavily on synergetics, synchronization and entrainment (Stepp and Turvey 2010; Stepp et al. 2011; Haken et al. 1985), all presupposing coupled (rather than fused) dynamical systems. If this is the only kind of boundary that stems from the free-energy principle, then there seems to be nothing in the idea of probabilistic inference per se that challenges enactive cognitive science (Bruineberg and Rietveld 2014; Clark 2015a). In other words, if the role played by the concept of inference in FEP is no more cognitively demanding than generalized synchrony, then it provides no grounds for distinguishing radical enactive approaches to cognition from those that are based on FEP.35

One might object that there is an important difference between simple life-forms or primordial soup simulations and more complex systems that predictive-coding is seeking to model. The behavior of the first class of systems might be understood in terms of dynamic coupling and reducing disattunement. The second class of systems necessitates the explicit computation of prediction-error based on a generative model and requires a more epistemic understanding of the internal workings of the animal: wherever there are systems that can select actions so as to deal with the absent or lie in the future, this requires explicit representational knowledge structures about these distal affairs, one might think (Seth 2013; Clark, personal communication).

However, as Orlandi (2015) points out as well, it is not clear that the structure of the generative model should be thought of as representational. We follow ecological psychology in arguing that the environment is rich with information that the perceptual system is able to directly access, because of our evolutionary history and the abilities and skills we learn. People and other animals (for instance primates) are situated in a very rich landscape of affordances (Rietveld and Kiverstein 2014). We proposed that the generative model is best thought of as a dynamical system of (affordance related) states of action readiness that reflect36 the hierarchical and temporal organization of the changing environment. As the animal develops skills, the generative model becomes more and more sensitive to the relevant particularities of the situation, and opens the animal up to the relevant affordances available in the environment.37

Some of these abilities and skills for engaging with the world may be reused in ways that may explain our capacities for planning, imagining and reasoning about counterfactuals (Anderson 2014). Other abilities for planning and imagining might come from the ways in which we actively structure the environment to scaffold our thinking (Rietveld and Kiverstein 2014). The exact balance of environmental and agential components is an open empirical question to be settled on a case-by-case basis (see the ethnographic description of architects at work in Rietveld and Brouwers (2016)). However, we believe that the richness of the landscape of affordances in which human embodied minds are situated will be an important part of the answer.

We wish to make three final points before we conclude. First, the explicit computation of a prediction-error will be based on structures that represent and encode long term information about causal regularities in the environment. We have argued that for many cases of free-energy minmization, it is unnecessary to make appeal to this type of structures within the organism. We leave it as an open question when, and if at all, it is necessary to explain the internal structure of the organism by appeal to representations. For our purposes we prefer to think in terms of affordance-related states of action readiness also for understanding forms of what is traditionally called ‘higher’ cognition. Second, the free-energy principle provides a clear functional continuity between simple living systems and complex ones like humans. Predictive-coding theories understood in abstraction from the free-energy principle run the risk of missing this important functional continuity, understood by us in terms of the tendency towards an optimal grip. Third, one of the big questions for future work in neuroscience is how more locally mechanistic accounts of brain functioning, like predictive-coding, can be integrated with a more broadly complex systems perspective on the brain. Phenomena like entrainment and synchronization seem to play a crucial role in brain functioning (Varela et al. 2001; Engel et al. 2001). We think the “ecological-enactive” reading of the free-energy principle provides a necessary framework for carrying out this integrative work. This is something one would fail to accomplish if one would exclusively focus on predictive-coding and ignore the free-energy principle.

The Helmholtzian interpretation of the anticipating brain brings with it a problematic conceptualization of the boundary separating the organism from the environment. The environment becomes a hidden cause that must be inferred from ambiguous sensory input, which is all the brain has access to. However we’ve argued above that there is no reason to accept such a Helmholtzian interpretation of the relation between the organism and its environment. Moreover, such an interpretation is in deep tension with the free energy principle according to which the brain is a part of a larger coupled system that on the basis of its coupling is constantly reducing disattunement with the environment. Through the organism’s minimization of free-energy, the brain’s internal dynamics are normally adequately attuned to the external dynamics of the environment.

Central to the predictive processing framework is the notion of a generative model. On the Helmholtzian interpretation, a generative model is a model that captures the causal structure of the agent’s environment. The generative model is then used to generate ‘mock’ sensory input that is then compared to actual sensory input to compute prediction-errors. Such an interpretation introduces a functional asymmetry between (top-down) predictions and (bottom-up) prediction-errors. Top-down predictions are produced on the basis of internally realized knowledge structures (Clark 2016). This may raise the worry that our ecological-enactive interpretation fails to accommodate this functional asymmetry.38 Without these internally realized knowledge structures, we just have complex looping interactions between brain, body and environment that mysteriously reduces disattunement.

On our interpretation, the function of the generative model is not to provide the agent with a representation of the dynamical structure of the environment per se, but rather to steer its interactions (over multiple timescales) with its environment in such a way that a robust brain-body-environment system is maintained. The organism’s internal structure and organization is then understood as multiple simultaneous and coupled affordance-related states of action-readiness that together shape (through top-down precision-modulation) the salience of solicitations in the environment. The self-organization of these states of action-readiness allows the animal to tend towards an optimal grip on the multiple relevant affordances in the situation. There is an asymmetry between the sensory inputs that flow from the environment and the anticipations of sensory input based on the organism’s history of interactions with the environment. However, we have been arguing that the internal dynamics can be understood in terms of patterns of action-readiness that constitute selective openness to multiple affordances simultaneously rather than from internalized knowledge structures. As we have proposed in earlier work (Bruineberg and Rietveld), the tendency towards an optimal grip on multiple affordances simultaneously can be explained as a metastable attunement to environmental dynamics. This metastable attunement allows for rapid and flexible switching between relevant action possibilities (Kelso 2012; Rietveld 2008b; Bruineberg and Rietveld 2014). Crucially, once we dispense with misleading/distracting talk of probabilistic inference, it is no longer necessary to understand past experience and learning as encoded in the form of representational knowledge structures.

Instead we understand past learning and experience as manifest in the skilled animal’s anticipatory dynamics to act in ways that improve grip on the affordances on offer in the situation.

Grip, as so conceived, is certainly not “the new representation” as Andy Clark has suggested (Clark 2015a). On Clark’s view, the generative model is estimating the current state of the organism in its environment, something that is only possible because of the knowledge the agent has of the world. On our view of the generative model, it is preparing the agent to perform actions that improve grip on multiple simultaneously relevant affordances in the situation. This is something that is only possible because of metastable attunement to environmental dynamics that make the agent open and ready for dealing effectively and flexibly with what the future may offer.