Communication as Socially Extended Active Inference: An Ecological Approach to Communicative Behavior

Abstract

In this paper, we introduce an ecological account of communication according to which acts of communication are active inferences achieved by affecting the behavior of a target organism via the modification of its field of affordances. Constraining a target organism’s behavior constitutes a mechanism of socially extended active inference, allowing organisms to proactively regulate their inner states through the behavior of other organisms. In this general conception of communication, the type of cooperative communication characteristic of human communicative interaction is a way of constraining interaction dynamics toward the goals of a given joint action by constructing and altering shared fields of affordances. This account embraces a pragmatist view according to which communication is a form of action aiming to influence the behavior of a target, and stands against the traditional transmission view according to which communication fundamentally serves to convey information. Understanding acts of communication as active inference under an ecological interpretation allows us to link communicative and ultimately linguistic behavior to the biological imperative of minimizing free energy and to emphasize the action-oriented nature of communicative interaction.

Acknowledgments

This research was produced thanks in part to Remi Tison’s Joseph-Armand Bombardier Canada Doctoral Scholarship (767-2019-1371). Pierre Poirier also wishes to acknowledge funding from SSHRC grant 435-2020-0846.

Notes

1 The notion of information at play here entails that of content. It should therefore not be confused with ecological information or with uncertainty-reducing information as in Shannon’s mathematical theory of communication (see note 4). Relatedly, the account of communication proposed in this paper targets the communicative behavior of organisms and is therefore not intended to replace Shannon’s theory of communication.

2 This is not the place to produce a complete list of references in all these domains for such a widespread view. Here is nonetheless a non exhaustive list of representative works either defending or characteristic of the transmission view: Bradbury & Vehrencamp, 2011; Corballis, 2017; Fitch, 2010; Fodor, 1975; Green & Marler, 1979; Hauser, 1996; Hinde, 1981; Horn & McGregor, 2013; Lewis, 1969; Loar, 1981; Lyons, 1977; Mann, 2018; Otte, 1974; Pinker & Bloom, 1990; Planer & Godfrey-Smith, 2020; Seyfarth et al., 1980, 2010; Stalnaker, 2014; Smith, 1980; Wiley, 2013).

3 See for example the following passages in the Essay concerning human understanding: “[words] being immediately the Signs of Mens Ideas; and, by that means, the Instruments whereby Men communicate their Conceptions, and express to one another their thoughts and Imaginations, they have within their own Breasts” (Locke, 1689/1975, III, iii, p. 6), or “the end of Speech is, that those Sounds, or Marks, may make known his Ideas to the Hearer” (III, ii, p. 2).

4 Although Shannon and Weaver’s work on communication has been taken as providing a bedrock for TV, it is not perfectly clear that they themselves would have endorsed it or its implications for a general conception of language, especially the identification of transmitter and receiver with speaker and hearer, since the former merely perform encoding and decoding functions (Shannon, 1948). They were also eager to distinguish their notion of information from traditional notions of meaning (Shannon & Weaver, 1949, p. 8; but see Gallistel, 2020).

5 One of its original proponents has even called it a “dismal failure” (Lycan, 2009, p. 551).

6 This is a simplified version of the classical “disjunction problem” (Fodor, 1984) for information-theoretic and teleosemantic theories of content. Another similar problem is the “distality problem”, where more or less proximal objects or states of affair in the causal chain leading to the instantiation of a representation fulfil the conditions of the theory but intuitively do not constitute the content of the representation. For instance, it is not clear why we should say, under standard informational or teleosemantic theories of content, that the neural activation of the frog should represent flies rather than the state of its retinal cells when it perceives a fly, which is also systematically correlated with this neural activation (see Roche & Sober, 2019 for a recent discussion of these issues, and also Artiga and Sebastian, 2020).

7 To be clear, we are not advocating an eliminativist stance according to which intentional ascriptions are never justified (see e.g., Churchland, 1981). Rather, we argue that we should not use such a vocabulary as an explanans in a scientific explanation of communicative behavior.

8 See also the “biogenic” approach to cognitive evolution (Barrett, 2018; Lyons, 2006), which states that cognitive processes should be seen as elaborations of basic biological processes. This view is opposed to the “anthropogenic” approach, which starts from cognitive categories inherited from the cognitive revolution (reasoning, theory of mind, natural language, memory, etc.) and attempts to find evolutionary histories accounting for them.

9 As we will see, these negative feedback control loops are just another description for the free energy minimizing activity of organisms described by the active inference framework that will be introduced in section 2.

10 The notion of surprisal must be distinguished from the notion of the psychological state of surprise. Surprisal is a measure of the distance between the probability distribution specifying the viable states for the organism and the probability distribution over the states that the organism finds itself in. A state with a high surprisal will often generate a psychological state of surprise. For example, I might be surprised to find that I’m falling from a ten story building, which is also a state of high surprisal for the kind of organism that I am, but those notions are clearly distinct.

11 This is an important lesson of the “crooked scientist argument” put forward by Bruineberg et al. (2018). An organism does not ultimately minimize free energy in order to establish an accurate representation of its environment in the way of a scientist establishing a theory, but rather to maintain its organisation. This is because it would be useless for an organism to simply accurately represent the fact that it finds itself in a non viable state if it does not also perform active inference to return to a viable state. If the organism is a scientist, it is a crooked scientist interested in establishing the theory that it exists.

12 We can think here of Bruineberg’s example of an academic researcher neglecting its metabolic needs by skipping dinner in order to finish its paper (2017, p. 9), or, more dramatically, an activist prepared to sacrifice her life to fulfil some political ideal.

13 Allen and Friston (2018) describe it as a “state theory,” compatible with many different “process theories” specifying the empirical details explaining the processes and mechanisms by which organisms manage to minimize free energy.

14 Indeed, predictive processing theory has been strongly inspired by the work of thinkers such as Helmoltz (1860/1962) and (Gregory, 1980), who viewed perception as a process of inferring hidden states of the world. This view of perception has strong affinities with classical cognitivism. However, although we cannot show that here, we believe that the most interesting elements of the predictive processing story can be retained and reinterpreted in a broadly ecological framework (see e.g., Kirchhoff & Robertson, 2018; Orlandi, 2016, 2018).

15 Given the recent debates concerning the compatibility of ecological psychology and enactivism (see e.g., Heft, 2020; Read & Szokolszky, 2020), as well as the fact that this paper do not draw on the specifically enactivist elements of Bruineberg et al.’s interpretation, we chose to name the interpretation of active inference proposed here the “ecological” interpretation of active inference, rather than the “ecological-enactive” interpretation.

16 We adopt here the relational construal of the notion of affordance (Chemero, 2003, 2009) over its dispositional alternative (Scarantino, 2003; Turvey, 1992), but this choice has no implications on the current issue.

17 This definition of the landscape of affordances is significantly different from its usual definition in the litterature, which appeals to Wittgenstein’s notion of a “form of life” (e.g., Bruineberg & Rietveld, 2014; Kiverstein et al., 2021; Ramstead et al., 2016). The added value of this notion is unclear to us, so we prefer more simply to define a landscape of affordances as a function of the action capabilities of an agent or a set of agents.

18 Given the schematic and introductory nature of this section, we will remain neutral on the numerous debates concerning the notion of ecological information in ecological psychology, such as whether there is information about affordances in the environment independently of its use by particular organisms (Baggs & Chemero, 2018; Hutto & Myin, 2017; van Dijk et al., 2015 ).

19 The notion of function is necessary to formulate this definition because it provides a way to distinguish signals, which aim to influence behavior, from cues, which influence behavior but without having the function to do so (Hauser, 1996; Rhebergen et al., 2015; Roberts et al., 2007). For example, mating calls of tungara frogs are signals for female frogs, but are cues for bats, who are their predator. We adopt a systemic or organisational account of the notion of function (Frick et al., 2019; Mossio et al., 2009; Saborido et al., 2011) against the traditional etiological account (Godfrey-Smith, 1994; Millikan, 1989; Neander, 1991), which generates problems for influence-based accounts of communication (see Frick et al., 2019).

20 The original quote states that “war is nothing but a continuation of politics with the admixture of other means.”

21 Vygotsky’s notorious analysis of the ontogenesis of the pointing gesture, analogous to the example given here, is an illuminating case for such a process (1978, p. 56). Vygotsky suggested that the pointing gesture of an infant originates from its behavior of reaching for a given object. Caregivers in its environment react to this behavior by giving the object to the infant. The infant therefore learns that the behavior of reaching for an object in the presence of its caregiver affects their behavior in such a way. She can then use this behavior (trying to reach, and ultimately just pointing) as an act of communication, to affect their behavior in order to minimize her free energy (eg. obtain an object that she desperately wants). In this case, the behavior of reaching for an object produces a solicitation of giving the object to the infant in the caregiver’s field of affordances because they have learned that an infant reaching for an object wants that object and because tending to the infant’s needs minimizes their own free energy.

22 If the behavior driven by the signal is systematically disadvantageous for the target, there will be some selection pressures for perceptual systems in which the signal does not solicit the behavior in question. It therefore seems that, for a given communicative behavior to become an “evolutionary stable strategy,” and so be maintained in time, acting on the solicitation created by the signal must minimize free energy for the target a statistically significant amount of time. In simpler terms, there has to be something in it for the target organism, otherwise it will just stop paying attention. This point is often made by proponents of the information-based approach to animal communication (see eg. Planer and Godfrey, 2020). However, proponents of influence-based views of animal communication have argued that communicative behavior that is advantageous for the signaler, and thus activate solicitations that do not necessarily lead to free energy minimization in the target, can be maintained over time : “an equilibrium point at which perceivers are somewhat disadvantaged is therefore probably common” (Owren et al., 2010).

23 It remains that, on our view, there will be some degree of continuity between acting on the field of affordances and acting directly on the behavior of others. Borderline cases stretching our intuitions about communication are therefore to be expected.

24 In its analysis of these cases of communication, TV is stuck in debates concerning the “honesty” of such signals, that is, the degree to which they are correlated with, and therefore truthfully indicate, aggressive motivation and behavior (see e.g., Hauser, 1996, pp. 450–471 and Laidre & Vehrencamp, 2008). These problems are completely evacuated in our account. Cases where organisms produce displays of aggression but does not follow it up with aggressive behavior are not to be understood as dishonest signalling, but rather as cases where either the act of communication achieved its goal and prediction error was reduced, or it failed and the organism, for various possible reasons, opted not to pursue this goal by other means (e.g., aggressive behavior).

25 This can be seen as a particular case of the traditional discussion concerning altruism and inclusive fitness in evolutionary biology (Axelrod, 1984; Birch, 2017; Hamilton, 1964; Skyrms, 1996; Sober & Wilson, 1998). Without delving into this complex literature and the evolutionary story of how this comes about, we can simply point out that, in the active inference framework, the production of altruistic behavior in social organisms could be explained by the instantiation of the relevant altruistic priors in their embodied generative models.

26 Bratman’s analysis includes a third clause according to which there is a common knowledge of the first two clauses among the participants of the joint action, entailing minimally that they know that the other knows that they share a common goal. This clause dramatically inflates the cognitive conditions required for participating in a joint action. Blomberg (2016) has argued persuasively that this third clause is unnecessary.

27 Gibson himself already considered the possibility that social cognition involves the perception of the affordances of others: “the child begins, no doubt, by perceiving the affordances of things for her (. . .). But she must learn to perceive the affordances of things for other observers as well as herself (Gibson, 1979, p. 141)

28 For instance, it is sometimes noted in the Gricean tradition (e.g., Schiffer, 1972) that informative acts of language, used to constitute the common ground of a conversation, aim to bring about beliefs that are “activated,” that is, present to the mind of the audience. Talk of “activated” or “unactivated” beliefs is an awkward and artificial addition to our usual folk-psychology, designed to supplement its limitations concerning the temporality of cognition and the cognitive limitations of agents. Our view, in which an “activated” belief can be understood as the simulation of an affordance or a field of affordances, provides a better framework to account for these kinds of spatiotemporal cognitive dynamics.

29 Of course, some common goals or contexts might confine interlocutors to more rigidly active or passive roles. For example, a speaker telling a story will probably take a more active role than the hearer in the interaction. But even then, the hearer will often engage in active listening and give some form of listener feedback (Bangerter & Clark, 2003; Duncan, 1974) that will be monitored by the speaker to regulate her ongoing communicative activity.

30 The distinction between shared sets of expectations constituting “local” acts of communication and shared sets of expectations constituting conventional acts of communication might be understood to be loosely mapping the traditional distinction between speaker meaning and linguistic or literal meaning (Allan, 1986; Carston, 2008; Grice, 1957; Recanati, 2002).

31 Interestingly, in the many studies attempting to learn symbol systems to apes, the acts of communication produced by the subjects were mostly of the imperative type. For instance, 96% of the spontaneously produced acts of communication of Kanzi the bonobo were requests (Fitch, 2010; Greenfield & Savage-Rumbaugh, 1990). This is consistent with the idea implicit in our framework that the “paradigmatic” act of communication is imperative rather than informative.

32 It is important to stress that simulations of fields of affordances are not themselves affordances. As remarked by de Wit et al. (2017), following its introduction in neuroscientific models, the notion of affordance has sometimes been incorrectly used to designate the patterns of neural activation making an organism responsive to an action possibility in its environment, rather than the action possibility itself. We wish to avoid such a mistake, leading to a slippery internalist slope, while still allowing the possibility for some organisms to simulate fields of affordances in order to coordinate their behavior with respect to absent or future affordances.

33 Glenberg et al. (2009) explicitly interpret such results as showing that linguistic comprehension of an utterance relies on the activation of the relevant affordances linked to the objects referred to by the utterance.

34 Another important component of linguistic communication neglected in this paper is syntactic structure. Our proposition is compatible with many current accounts (for example, usage-based (Christiansen & MacDonald, 2009; Lieven, 2016; Tomasello, 2003) or functional (Givón, 1995, 2014) accounts of syntactic structure), but particularly promising links could be developed between our account and views according to which syntactic structure emerge at least in part from general constraints on motor sequencing applying to action in general (Casado et al., 2018; Knott, 2012; Koranda et al., 2020; Lashley, 1951).

35 See Egan (2014, 2020) for interesting work on the role of intentional descriptions in cognitive science and what she calls the “intentional gloss.”