Predictive Coding: Towards Explaining the Evolution of Theory of Mind

Introduction 

Unlike most species, humans possess sophisticated minds with unique social capacities such as Theory of Mind (TOM). Cognitive neuroscientists have begun to identify the underlying neural mechanisms of TOM, which has led to the recent development of a Predictive Coding (PC) approach. In what follows, I shall define human TOM and PC. From there, I will outline a traditional model that argues for the importance of sociocognitive prediction in the evolution of TOM. At this point, select recent fMRI evidence supporting the PC account of TOM will be introduced. Limitations and questions left unanswered by the PC account will then be considered. Finally, it will become clear that the Predictive Coding framework may open an important window to explaining the evolutionary path and neural computations underlying Theory of Mind.

Theory of Mind

I will begin with a definition of human TOM. To interact effectively in social settings we need to understand and interpret others, which involves making educated guesses (or predictions), as to their emotional and cognitive mental states. Crucially though, we’re restricted to the external and observable side of others – we don’t have immediate access to the unobservable causal structure behind their actions. We don’t have direct access to a person’s beliefs, desires, intentions, goals and preferences. Given that restriction, the ability to reason about the potential causes of others’ actions is a fundamental requirement for social cognition. It’s what allows us to gauge if our friend keeps looking at her phone because she needs to keep track of the time or because she’s bored. It helps us judge if our waiter is being intentionally rude or is just unusually stressed. This ability to reason in an attempt to determine the emotional and cognitive states of others is known as Theory of Mind (Koster-Hale & Saxe, 2013). In the next section I will summarise what we know about how TOM develops and which parts of the brain are implemented.

Development and Brain Regions

There is general agreement that TOM largely develops from our own early motor and perceptual experiences and is strengthened by expectations we internalise in childhood from early social encounters (Flinn, Geary & Ward, 2005; Koster-Hale & Saxe, 2013). These expectations then become predictions about the cognitive states and resultant actions of other organisms (Brown & Brune, 2012; Koster-Hale & Saxe, 2013). Neuroimaging studies in recent years point to the notion that TOM “seems to depend on a distinct and reliable group of brain regions, sometimes called the “mentalizing network…which includes regions in human superior temporal sulcus (STS), temporo-parietal junction (TPJ), medial precuneus (PC), and medial prefrontal cortex (MPFC)” (Koster-Hale & Saxe, 2013, p.836). In (2000), Frith and Frith concluded that the most notable brain areas activated during TOM tasks are the medial prefrontal cortex (MPFC) and temporo-parietal junction (TPJ). Since then, over 400 studies have been published on these regions (Koster-Hale & Saxe, 2013, p.836). While there is widespread agreement on which parts of the brain are activated during TOM processes, much less is known about the neural representations and predictive computations implemented in these regions (Brown & Brune, 2012; Koster-Hale & Saxe, 2013). Some cognitive neuroscientists aim to bridge this explanatory gap by incorporating a theoretical framework known as predictive coding (PC), which I shall now explain.

Predictive Coding 

PC is a framework within cognitive neuroscience that aims to provide a unifying explanation of perception, learning and action (Friston, 2005; Hohwy, 2013; Clark, 2016). The central idea is that neural systems approximate Bayes Theorem to produce forward-looking predictions generated by an organism’s prior model of the world. These predictions cascade in a top-down fashion to inform expectations about incoming information. If an organism encounters sensory stimuli it did not predict, a prediction error (PE) will be generated. These bottom-up (PE) signals meet top-down predictions at each successive level of the cortical hierarchy. They move both laterally and upwards through the hierarchy to update internal models of reality and more accurately predict future incoming information (Hohwy, 2013; Clark, 2016).

The driving goal behind the framework is to minimise prediction error and generate accurate future predictions about the state of the organism and the world. This is important because organisms are subject to an ongoing barrage of sensory stimuli. On the PC view, judging which of those signals are relevant to survival as well as predicting the cause of those signals is necessary to maximise chances of survival (Friston, 2010; Clark, 2016). In recent years, the PC framework has been swiftly employed as an explanatory tool across multiple disciplines and to a diverse array of phenomena, from motor action and sensory perception to schizophrenia and autism. While the field is in its infancy, it is well accepted that the visual system processes stimuli in accordance with the PC framework. A review conducted by Chanes and Barrett (2016), demonstrates that the PC approach is increasingly supported by functional, experimental and anatomical evidence (for a detailed review see Chanes & Barrett, 2016). The focus of this paper, however, is on the relevant evidence and potential of the PC framework as it relates to TOM. With an understanding of PC in hand, we’re now in a position to examine an existing model that argues for the relevance of prediction in the evolution of social brains.  

The Importance of Prediction for Evolving Social Brains

In this section I will refer to the work of Alexander (1989; 1990) and Flinn, Geary and Ward (2005). In particular I shall focus on the Ecological Dominance–social Competition (EDSC) Model conceptualised by Alexander (1989), which was developed and investigated further by Flinn et al. (2005). The puzzle these authors tried to solve was why human beings developed such advanced social cognitive capacities in contrast to our closest biological relatives. Flinn et al. (2005) argue that traditional hypotheses centred on ecological demands such as hunting, tool use and foraging are unsatisfying when it comes to explaining the human evolution of extraordinary cognitive traits such as TOM. These traditional ecological constraints, they suggest, may have been more of a secondary source of recent cognitive evolution, compared with the significance of psychological adaptations, which function to contend with complex social relationships (Flinn et al. 2005, p.13).

The authors draw from work by Alexander (1989; 1990) to propose a comprehensive and integrated explanation for the development of a range of human cognitive capacities including TOM. Once humans had mastered the hostile, challenging forces of nature, they argue, selective pressures resulting from conspecific competition became increasingly important (Flinn et al. 2005). As they note, social relationships are variable and complex. Social success requires the ability to predict the future moves of others in order to employ countermoves, whether dealing with competitors or cooperators. This difficulty is amplified by the development and shifting of social coalitions and the requirement to manage networks formed by multiple relationships. 

In line with this, Alexander (1990) argues that the sheer variety of human social puzzles is seemingly infinite; no two individuals ever appear to be in the exact same social environment and it seems that no two social situations could be precisely identical. He also notes that human social relationships can change swiftly, requiring rapid modifications in strategy. In this way, Alexander (1990) suggests that individuals with the capacity to mentally rehearse or construct possible social scenarios would have an increased ability to contend with a range of novel social situations, as well as forces like cultural change. In his view, once our ancestors had sufficiently mastered their ecological environments, social cleverness, including capacities like TOM would have become paramount. Specifically, he suggests that “unlike static ecological challenges, the hominin social environment became an autocatalytic process, ratcheting up the selective advantage associated with the ability to anticipate the social strategies of other hominins and to mentally simulate and evaluate potential counterstrategies (Alexander, 1989, p.469).

With the Ecological Dominance–social Competition (EDSC) Model, Flinn et al. (2005) forwarded Alexander’s (1989; 1990) suggestion that the importance of social relationships and coalitions led to inter-species competition. They agree that the successes of individuals and coalitions would have depended in a significant way on sociocognitive capacities such as TOM (which crucially includes an ability to predict incoming sensory information). This, they contend, would have required the development of expanded neurological structures (Flinn et al. 2005, p.15). They conducted investigations into two empirical sources: (1) the hominin fossil record, and (2) human neurobiology and cognition. A thorough explanation of their methods and results falls outside the scope of this paper, but for our purposes a summary of their conclusion shall suffice. The authors found that the EDSC model was supported by “a unique combination of coevolved characteristics and their temporal sequencing” (Flinn et al. 2005, p.35). These included unusual patterns of speciation, extended childhood stages (which they suggest may be necessary to acquire complex social skills), a neocortex arguably larger than primates in the areas that support TOM and a convergence with species which also compete among socially complex coalitions, such as chimpanzees and dolphins (see Flinn et al. 2005, p.35-37 for an extended description of findings).

The work developed by Alexander (1989; 1990) and Flinn et al. (2005) emphasises the importance of the ability to predict social scenarios and offers an explanation for the development of neurological structures that could support sociocognitive abilities such as TOM. It therefore appears that this work provides a potential fit with the Predictive Coding approach to the evolution of TOM. Now these accounts have been outlined I will turn to recent empirical evidence supporting the PC approach to TOM.  

Empirical Evidence for the Predictive Account of TOM

In recent years, the predictive approach has been applied to work in the field of social cognition. This compliments existing research; as noted by Brown and Brune: “there is already a large body of empirical and theoretical neuroscientific work from neurophysiological, behavioral, and computational perspectives that provide substantial evidence for a fundamental role for predictive mechanisms in the processing of social information and in social interaction” (Brown & Brune, 2012, p.1). Given this evidence already links predictive mechanisms with the processing of social information and social interaction, it seems reasonable to suggest that PC may play a stronger explanatory role when it comes to TOM specifically. A review of evidence supporting the “central role of predictive mechanisms and Bayesian inference in social interactive processes” can also be found in Brown and Brune (2012, p.1). For the sake of brevity, however, I will now focus on a particular fMRI review that supports the PC approach to TOM. 

Predictive Coding Signatures at Three Levels

Koster-Hale and Saxe (2013) reviewed fMRI evidence on neural responses during specific interpersonal scenarios at three levels: (1) to goal-directed action and biological motion in the superior temporal sulcus (STS), (2) to other’s desires and beliefs in the temporo-parietal junction (TPJ), and (3) to others’ stable personality traits in the medial pre-frontal cortex (MPFC). Recall that these areas of the brain are well accepted as those activated during TOM cognitive processes. These three regions also roughly correspond to proposed temporal scales represented by levels in the cortical hierarchy in the PC model: moment-by-moment predictions reside at the lowest levels, intermediate predictions sit higher, and stable ongoing beliefs reside at the highest levels of the hierarchy (Hohwy, 2013; Clark, 2016). On this view applied to TOM, we predict and interpret in-the-moment biological motion like eye movements. We also infer and predict mental states across longer periods of time – we might, for instance, predict that because your goal is to finish a paper tonight, you’d rather not accompany us to a party. At even longer timescales, being socially successful depends on predicting with a sufficient degree of accuracy things like the reliability of your spouse or the trustworthiness of your employee. 

Koster-Hale and Saxe (2013) reviewed a wide range of fMRI imaging studies that measured neural responses in the aforementioned three brain regions corresponding to TOM activity. Consistent with the PC framework, they found that unexpected stimuli systematically elicited higher activity with respect to each region’s level of preference and temporal abstraction than expected stimuli. The significance of this finding is that reduced response to expected stimuli is a key signature of PC; therefore these results across multiple studies are what we would expect to see if PC is on the right track. The authors concluded that as this signature was demonstrable in each of the three brain areas widely believed to correspond with TOM, these results strengthen the notion that TOM can, at least to a certain extent, be explained by reference to the PC framework. For a detailed discussion of the reviewed studies and findings see Koster-Hale and Saxe (2013, p. 839-844).

Limitations

Of course, there are limitations with the empirical evidence gathered to date and objections one could mount against the suggestion that the Ecological Dominance–social Competition Model provides a compelling account of the evolution of TOM. I will now consider each of these, starting with the EDSC Model. 

As Flinn et al. (2005) note themselves, one of the biggest challenges facing this account is that it’s practically impossible to test or falsify it. This is because, to date, our best empirical and theoretical investigations into the evolution of cognitive traits like TOM are inconclusive. We’re not certain about the environmental factors that may have contributed; we don’t have sufficiently close biological relatives to compare ourselves with, and thus theories like this seem necessarily speculative in nature. This doesn’t entail, however, that Alexander (1989; 1990) and Flinn et al. (2005) are wrong. To the contrary, I argue that their theoretical arguments are detailed, thorough and compelling. They undertook empirical investigation into the hominin fossil record and human neurobiology and cognition to buttress the appeal of their conceptual arguments. While we can’t be certain they are on the right track, if future research into TOM and PC aligns with their account it may serve to fortify it. 

As for the fMRI evidence for a PC explanation of TOM, Koster-Hale and Saxe (2013) are similarly forthright when it comes to limitations with their review. Demonstrating a key signature of PC across three regions of the brain widely believed to be responsible for predicting (1) goal-directed action, (2) beliefs and desires, and (3), preferences and personalities is a good start. They conclude, however, that as the field is in its infancy, more investigation is required to strengthen the claim that PC plays a crucial explanatory role in TOM. They suggest three important areas of research that could serve to solidify this claim. These involve breaking down (1) what constitutes a prediction error, (2), what kinds of predictions each brain region makes, and (3), what kind of information directs those predictions. If we can work towards these aims, they contend, it might be feasible to “formulate much more specific hypotheses about the computations, and information flow, that underlie human theory of mind.” (Koster-Hale & Saxe, 2013, p.844)

Concluding Remarks

In accordance with the work outlined above, it’s legitimate to suggest that the evolution of human sociocognitive processes such as TOM emerged through the cooptation of predictive mechanisms integral to the PC framework, or vice versa. It’s plausible that selection pressures on sophisticated abilities to predict the behaviour of others would have increased in early hominid environments if ecological challenges became less significant than social relationships, coalitions and inter-species competition. 

The purpose of this paper was to examine relevant evidence gathered to date in favour of the claim that the PC framework can help explain the evolution and underlying neural mechanisms of TOM. In sum, the EDSC model, while vulnerable to charges of being speculative and untestable, offers a compelling argument including explanations for extended childhood stages (which may be necessary to acquire complex social skills), a neocortex arguably larger than primates in the areas that support TOM and a convergence with species which also compete among socially complex coalitions.

In addition, there is extensive evidence of a key signature of predictive coding in fMRI studies of TOM: reduced responses to expected stimuli. This specific focus on TOM builds on a body of neuroscientific work that already provides substantial evidence for a fundamental role for predictive mechanisms in social interactions and in the processing of social information. Future research designed to more directly test how predictions and errors are constituted, directed and represented in different brain regions promises a twofold effect: it could provide support for models like the EDSC and open an important new window to explaining the evolutionary pathway and neural computations underlying Theory of Mind.

References

Alexander, R. D. (1989). Evolution of the human psyche. In P. Mellars, & C. Stringer (Eds.), The human revolution: behavioural and biological perspectives on the origins of modern humans. Princeton: Princeton University Press.455–513

Alexander, R., & University of Michigan. Museum of Zoology. (1990). How did humans evolve? : Reflections on the uniquely unique species. Ann Arbor, Mich: Museum of Zoology, University of Michigan.

Brown, E., & Brüne, M. (2012). Evolution of Social Predictive Brains? Frontiers in Psychology, 3, 414.

Chanes, L., & Barrett, L. F. (2016). Redefining the role of limbic areas in cortical processing. Trends in cognitive sciences, 20 (2), 96-106. 

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. The Behavioral and Brain Sciences, 36(3), 181-204.

Flinn, M., Geary, D., & Ward. C. (2005). Ecological dominance, social competition, and coalitionary arms races: Why humans evolved extraordinary intelligence. Evolution and Human Behavior, 26(1), 10-46.

Friston, K. (2005). A Theory of Cortical Responses. Philosophical Transactions: Biological Sciences, 360(1456), 815-836.Society of London B: Biological Sciences, 360(1456), 815-836.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127-38.

Frith, C.D., & Frith, U. (2000). The Physiological Basis of Theory of Mind. In Understanding Other Minds: Perspective From Developmental Soc Neurosci, S. Baron-Cohen, H. Tager-Flusberg, and D. Cohen, eds. Oxford: Oxford University Press. 335–356.

Hohwy, J. (2013). Predictive Mind. Oxford University Press.

Koster-Hale, & Saxe. (2013). Theory of Mind: A Neural Prediction Problem. Neuron, 79(5), 836-848. 

Manera, V., Becchio, C., Schouten, B., Bara. B.G., & Verfaillie, K. (2011). Communicative Interactions Improve Visual Detection of Biological Motion. PLoS ONE, 6(1), E14594.

Neri, P., Luu, J. Y., & Levi, D. M. (2006). Meaningful interactions can enhance visual discrimination of human agents. Nat. Neurosci. 9, 1186–1192.