Veissière et al. provide a compelling account of the use of variational principles (specifically, active inference) to provide a formal basis for understanding culture and cognition. This affords an opportunity to exploit the tools that come along with active inference in the context of social sciences. As an example of this, Veissière et al. highlight the importance of shared “regimes of attention,” and the way in which the elusive concept of attention may be pinned down in formal terms. This relies upon the concept of a Markov blanket (Pearl Reference Pearl and Smets1998) that statistically insulates the inside of a system from the outside. In this commentary, we discuss three perspectives on choosing a Markov blanket (Fig. 1). The first is the choice that we as scientists make when deciding upon our object of study. It is this that underwrites the application of Bayesian mechanics across interdisciplinary boundaries. The second is the implicit choice made by internal states of a Markov blanket as to which blanket states most influence their dynamics. This is an attentional process modulating influences from the outside in. The third is a choice between hypothetical blankets in a dynamical setting. It is this that determines how inside influences outside and gives rise to the concepts of salience and novelty (Clark Reference Clark2017a).

Figure 1.

Blankets, inference, and attention. This figure sets out the various ways in which we can “choose” a Markov blanket. The top image sets out the conditional dependencies between internal (μ), external (η), and blanket (b) states, where arrows show the direction of causation. Blanket states comprise active (a) and sensory (s) states. Given a mapping (σ) between the most likely internal and external states (μ and η, respectively) as a function of b, both internal and external states can be viewed as performing a gradient ascent on the same log-probability density. This is the non-equilibrium steady state density, or generative model. Technically, these dynamics maximize the evidence for the generative model, and are sometimes described as “self-evidencing” (Hohwy Reference Hohwy2016). The middle schematics show two alternative delineations of a Markov blanket in a social context. Either two brains sit within the same Markov blanket and can be thought of as jointly inferring their environment (η), but not each other, or they could be thought of as being on either side of a blanket. In this setting, each individual draws inferences about the other as the other is part of the environment. The bottom images set out the distinction between attention and salience from the perspective of a Markov blanket. The image on the left shows multiple sensory states (superscripted) and shows the form of the internal state dynamics (under Gaussian assumptions). The magnitude of the influence of each s on μ depends on the precision Π_s with which that sensory state depends on external states. The image on the right shows a different sort of selection, cartooning three alternative paths (indexed by i) that can be scored in terms of the expected log probability of the blanket states following that trajectory (where x[τ] means the path x follows over time). Note that the expression for the bound on this probability includes a relative entropy (the difference between the two H terms) that quantifies the salience or information gain expected along that trajectory. Interestingly, the entropy of blanket states conditioned on external states is inversely related to Π_s, highlighting the point of connection between attention and salience that often underwrites their conflation.

Selecting an object of study involves explicitly or implicitly segregating that thing from other things. This selection defines a Markov blanket that mediates interactions between that object and everything else. If we select the nervous system, the blanket comprises sensory receptors and muscles, whereas internal states are those neurons that respond to the former and drive changes in the latter. Because the system as a whole persists over time, it must be at (non-equilibrium) steady state. This implies dynamics that correct deviations from a steady state density, ensuring the system continues to occupy regions of high probability. Because the internal states of the system are coupled to external states, but only via the blanket states, it appears that internal states vicariously infer what is happening in the outside world (Friston Reference Friston2019). This Bayesian mechanical perspective says something intuitively sensible from the perspective of a nervous system: the brain draws inferences about the world based upon its sensory input.

We can frame the dynamics of any blanketed system in the same way. For example, if we take a more fine-grained approach we could treat an individual neuron as our system of interest (Palacios et al. Reference Palacios, Isomura, Parr and Friston2019), with its blanket comprising pre- and post-synaptic membrane potentials. The interesting thing about this is that the other neurons in the brain, previously internal states, have become external states. This means they have gone from performing inference to being inferred. Nothing has changed in the dynamics of the system itself, but by changing our perspective, we change the inference problem (i.e., generative or internal model) that is being solved.

This has two interesting consequences. The first is that it endorses the use of inferential formalisms at a range of scales (Kirchhoff et al. Reference Kirchhoff, Parr, Palacios, Friston and Kiverstein2018), whether cellular, cognitive, or cultural. The second is that the choice we make as to where the Markov blanket is drawn has consequences for how we think about the interactions between different parts of a system. Bringing this back to the question of cognition and culture, we could think of many individuals as the internal states of a system jointly inferring their shared environment or we could think of an individual drawing inferences about other individuals. In either setting, the challenge going forward is to set out the generative model from which inferential dynamics at a cultural scale emerge.

We now take the perspective from inside a blanket, and ask what it means to choose between alternative blanket states. This choice has two parts to it. The first is deciding which blanket states should influence internal state dynamics. The second is deciding between hypothetical trajectories the blanket states could follow. The distinction between these is formally identical to that between attention and salience attribution; two important features that emerge from solving a generative model.

Starting with attention, imagine we have multiple sensory states in a blanket. The degree to which each of these may be used to draw inferences depends upon the precision (inverse variance) with which they are predicted by external states, under the non-equilibrium steady state density. This manifests as a form of gain control, where those sensory states that are precisely predicted by external states are amplified relative to others in setting internal dynamics, exactly as in attentional gain control (Desimone Reference Desimone1996; Hillyard et al. Reference Hillyard, Vogel and Luck1998; Shipp Reference Shipp2016).

Attentional gain must be distinguished from the process of salience attribution (Parr & Friston Reference Friston2019). The latter involves overtly (Rizzolatti et al. Reference Rizzolatti, Riggio, Dascola and Umiltá1987) acting upon the world to obtain more information (Mirza et al. Reference Mirza, Adams, Mathys and Friston2016). This requires the capacity to score alternative trajectories (e.g., eye movements to different locations) in terms of their anticipated information gain (Lindley Reference Lindley1956). The relative probability for each trajectory is bounded by an expected free-energy functional. This functional favours those trajectories for which the salience is greatest (Parr et al. Reference Parr, Da Costa and Friston2020). As such, the process of salience attribution may be formalized as the process of choosing between alternative blanket trajectories.

Once a Markov blanket has been drawn around a system of interest, this licences an inferential interpretation of its dynamics. The choice of blanket tells us what is being inferred (external states) and what is doing the inferring (internal states). The advantage of appealing to a formalism of this sort is that it provides an opportunity to precisely define and simulate cognitive (and cultural) processes. We highlight the examples of attention and salience. These may be understood through the metaphor of a scientist who decides upon the quality of her data (i.e., attention) before drawing inferences, and then decides upon the next experiment to perform (i.e., salience) to optimize the quality of future data.