The formation of social beliefs and their role in orchestrating our social environment

The formation of social beliefs and their role in orchestrating our social environment

Matej Hrkalović Tiffany, 12292664 Literature thesis, 2020

Supervisors: Casper Hesp & Jan Engelmann MSc Brain and Cognitive Science Faculty of Science

1

Table of contents

Abstract ... 2

Introduction ... 3

1. Terminology: Beliefs, impressions, attitudes and expectations ... 5

2. Social perception and learning ... 6

2.1 Bayesian predictive brain hypothesis ... 6

2.2. Selecting socially relevant evidence ... 10

2.3. Priors involved in social inference ... 11

2.4. Biases involved in belief updating process ... 12

2.5. From traits to behavior to internal models ... 14

3. Hierarchical person-specific generative model ... 17

4. Person-specific priors and expectations modulate our actions and reality ... 19

5. Neural correlates of Bayesian active inference in social cognition... 22

6. Conclusion ... 24

2

Abstract

Ability to make higher-order inference about other people gives us the possibility to predict their thoughts and actions, reducing the experienced person-related uncertainty. Such inferences rely on internal models of one's social environment, including specific personality models and prior beliefs. However, little is known about how people construct, update and retain these person-specific beliefs. In order to better understand the steps of person-specific belief formation and belief updating, the aim of this review will be to present the „lifecycle“ of person-related beliefs and explain how they come together to affect our behavior and reality. Additionally, to put an emphasis on advantages of using interdisciplinary approaches, a Bayesian framework is going to be used.

3

Introduction

Imagine a situation where a colleague comes to you with a worried facial expression and asks you to lend him a substantial amount of money to cover his credit debt. In a split-second, many questions arise, and you start to evaluate your internal beliefs about the person. How reliable and trustworthy is he? Is he going to pay the money back? Ultimately, will you lend him the money? How confident are you in your choice? Such adaptability to our social environment depends on our ability to use observable events (e.g., behaviors and emotional expressions) to form useful inferences about higher-order (unobservable) variables (Koster-Hale & Saxe, 2013; Seo & Lee, 2012). This ability is what helps individuals to interpret, predict, and navigate their social environments. Such inferences rely on internal models of one's social environment, including specific personality models that encapsulate our prior beliefs (Brown & Brune, 2012; Brod, Werkle-Bergner, & Shing, 2013).

Social psychology and neuroscience researchers aimed to study the underlying processes of social cognition, but the precise mechanisms by which higher-level social inferences are linked to real-time action-perception cycles are still largely unclear. To capture this process in more detail, newer research began to use computational models (Corradi-Dell'Acqua et al., 2015; Chang & Koban 2013; Lind, Ghirlanda, & Enquist, 2019). Christian & Griffiths (2016) endorsed this trend by saying that using computer science and algorithmic thinking about the world can help us better understand the nature of our decision making or learning. Recently, a growing number of researchers attempted to account for social interaction using Bayesian models (e.g., Veissière et al., 2019; Bolis & Schilbach 2017), which give a solution to the problem by combining preexisting beliefs (top-down expectations) with obtained evidence (bottom-up evidence; Bach & Schenke, 2017; Hassabis et al. 2013; Macrae & Bodenhausen, 2000; Heleven & Van Overwalle, 2016). However, it remains to be specified how exactly we extract meaningful information about other people's traits and characteristics. Furthermore, we could ask how that information is maintained and interconnected to form consistent beliefs about other person's traits. Finally, closing the action-perception loop, we need to understand how exactly person-related beliefs come to have an impact on our action-selection towards other people and what role they play in orchestrating our interactions with others and the world. In this review, we aim to provide an integrative view of theoretical and empirical work on the formation and function of person-specific beliefs by examining the entire 'lifecycle', with a focus on how such beliefs give rise to expectations about our own actions and those of others.

4

Here, the process of belief formation will be thought of as a process of evidence accumulation where every interaction is seen as one iteration in which an individual can gather evidence about another person and update the probability distribution over a certain state, using the Bayesian approach (Christian & Griffiths, 2016; Friston et al., 2017). For that reason, belief formation and learning are going to be used as synonyms. Baker, Saxe & Tenenbaum (2011) argued that the Bayesian approach could be used to model inferences about mental states, in which behavior understanding is acquired by combining top-down expectations with bottom-up evidence. More specifically, individuals can use their person-specific prior beliefs and novel evidence to predict another person's behavior (Corradi-Dell'Acqua et al., 2015). For example, if we have a belief that another person is largely trustworthy (i.e. truthful and sympathetic to our well-being), we can use this belief as a basis to interpret, anticipate, and respond to a person's behavior. It is important to emphasize that we are considering subjective beliefs, held with varying degrees of certainty. By subjective belief, we mean that it does not have to resonate with reality, as people can (and often do) interpret and predict each other's behaviors based on oversimplified or even erroneous beliefs (Bach & Schenke, 2017).

Nowadays, the Bayesian approach has been applied across different levels of cognition, from lower-level (i.e. visual perception) to higher-level cognition (i.e. learning & metacognition; Moutoussis, Fearon, El-Deredy, Dolan & Friston, 2014). In this review, we will focus on higher-level cognition, where individuals use lower-level observations (i.e. actions and behavior) to make inferences about hidden states (personality or traits) which enables them to reduce uncertainty and create a more stable model of their social environment. In other words, learning is seen as a process of updating beliefs (i.e. model parameters) about another person, by integrating old and new information (Mathys, Daunizeau, Friston & Stephan, 2011).

5

1. Terminology: Beliefs, impressions, attitudes and expectations

Before going deeper into belief and representation formation, there is some variability in terminology and relations between different outcomes of impression formation, which need to be addressed. The topic of person-specific beliefs or knowledge formation is interdisciplinary, which means it is researched in different fields (i.e. psychology, sociology, evolution, neuroscience, and neuroeconomics) that put emphasis on different parts. Accordingly, the terminology of similar, or even identical, processes differs between research groups. For example, the bundle of information we use in order to make predictions and inferences about other people, within our social environment, have been referred to as person construal (Quinn & Rosenthal, 2012), person knowledge (Wagner, Haxby, & Heatherton, 2012), social cognition networks (Lind et al. 2019), and impressions (Park, 1986). Here, we use a common umbrella term – person-specific beliefs. In this integrative review, we synthesize literature on online active inference about other people's traits with literature on person-related construals, giving rise to a hierarchical generative model of beliefs about other's identity and personality traits. Traits will be thought of as typical patterns of emotions, behavior or ways of thinking that people usually engage in across different situations (Moutoussis et al., 2014).

Furthermore, we need to account for interrelated concepts of impressions, beliefs, assumptions, attitudes, and expectations. From a Bayesian perspective, beliefs about different states of the world or characteristics of the person can be seen as probability distributions over a group of possible options, which are updated by integrating evidence (Christian & Griffiths, 2016). Beliefs differ from attitudes, as they do not need to incorporate an evaluative component or preference, which is typical for attitudes (Connors & Halligan, 2015). However, there is a substantial overlap between them. Most importantly, beliefs also contribute to expectations, which interact and filter the incoming evidence in a top-down fashion. Expectations are based on what was previously stored in our memory (e.g. person-related construals and beliefs; LeBerge, 1995), and help us with the reduction of the probability space of alternative options. On the other hand, impressions include most of the previously named constructs. Impressions could be defined as a combination of person-specific beliefs with beliefs about how these traits relate to the observed behavior and the attitude towards that person (e.g. evaluative component) (Vonk, 1994). However, in order to understand how these different levels, inform each other, we need to understand the underlying process of belief formation.

6

2. Social perception and learning

2.1 Bayesian predictive brain hypothesis

In our everyday life, we gather evidence to inform our actions. However, such evidence is often incomplete and accompanied by varying degrees of uncertainty (Ting, Yu, Maloney & Wu, 2014). Meaning, our social beliefs, expectations, and decisions are based on inferences we make while considering previous experience and knowledge we have about causalities of events in the world. The Bayesian approach captures how such inferences are made by combining two sources of evidence. The first is our past experience (i.e., prior knowledge and beliefs) – summarized as probability distributions across hidden states of the world, model parameters, and model structure. The second is the likelihood of current observations, given hidden states of the world (Ting et al., 2014) – summarized as a matrix containing conditional probabilities for each observation, given each possible scenario. Bayes' rule then prescribes how to combine prior beliefs about hidden states of the world (s) with observable events (o) to arrive at posterior beliefs about the world (figure 1a):

𝑃(𝑠|𝑜) =

𝑃(𝑜|𝑠)𝑃(𝑠)

𝑃(𝑜)

Here, P(s|o) denotes the posterior confidence in each hypothesized state, given the data observed. P(o|s) denotes the likelihood of observing these outcomes under each possible state. P(s) and P(o) denote the prior probabilities of hidden states s and observations o, respectively. At the core of Bayesian inference lies the iterative process of belief updating (Moutoussis et al., 2014): combining prior beliefs and observable evidence to arrive at posterior beliefs, which then inform the prior beliefs for future observations (Figure 1).

7

Figure 1. Basic model of Bayesian inference of inferring which hidden states provide the optimal explanation for observed evidence. D presents a prior belief about hidden states P(s). S denotes hidden states of the world about which we have beliefs about and update them based on external information (O). A represents the likelihood matrix, where outcome expectation = A*s. This conditionality can also be shown in a graphical way, by using probability distributions. Based on the figure adopted from Yanagisawa and colleagues (2019) it is visible that we combine evidence with priors in order to get to the posterior. For instance, let's take an example in which we are trying to infer if our future colleague is trustworthy. Some of our good friends provided us the information about his good reputation (i.e. that he is known to be very trustworthy and reliable), we can use this information to inform our priors (orange probability distribution). However, as in most interactions we inevitable gather evidence from other people's behavior and we also gathered some evidence from the colleague’s behavior, we can generate a distribution of their likelihood (blue probability distribution). Now, having both probability distributions based on our priors and observable evidence, we can use them to update our current belief by computing them in posterior probability distribution (shown in green). However, if we do not receive information prior to getting to know the person, this distribution should resemble a uniform distribution. However, this example eliminates the dynamic characteristics of active inference in social interactions. As social interactions are more dynamic, there is a need for a broader model (b), which specifies how hidden states transition from one state to the next (by incorporating state transition probability matrices Bt). As the complexity of the model increases, belief updating

8

Mathematically working out equation (1) is called exact Bayesian inference. In practice, however, it is too computationally taxing. A generalized way to perform approximate Bayesian inference (in a way that is biologically plausible) is through the minimization of variational free energy, a measure of uncertainty or prediction error that quantifies the discrepancy between the observed and expected outcome (Friston & Stephan, 2007). According to the free energy principle (FEP), the processes underlying cognition and action can be understood as minimizing uncertainty (i.e. free energy or prediction error), which is equivalent to maximizing model evidence (Friston & Stephan, 2007; Parr & Friston, 2019). The FEP is a computational expression of the fact that individuals need to limit the possibilities of physiological and perceptual states in which they can find themselves, in order to function in their environment (e.g., needing a certain degree of social belonging).

During active inference, individuals are cast as implicitly employing a generative model to infer the most probable latent cause (Parr, Rees & Friston, 2018). The model reflects an individual's relationship with the environment and enables predictions about the observations, while its parameters project an individual's beliefs about its relationship with the environment (Figure 1). Internal states represented by the generative model also encode a probability distribution or prior (Bayesian) beliefs about causality represented in the current state. The generative model is used as a reference point in the circular action-perception process that is characteristic of active inference. When incorporating action in active inference, an extension of the previously shown generative model needs to be introduced (see Figure 2). Action (π) is cast as selecting between transition matrices (denoted B) that describe the temporal dynamics of the states of the world. A priori, an agent expects to select actions that minimize the expected free energy of the future. Expected free energy can be decomposed as the sum of the expected risk and ambiguity for each course of action. Where risk denotes the difference between the anticipated and preferred outcomes. Ambiguity is the uncertainty related to the contextual differences of the learning environment (Hesp, Smith, Allen, Friston & Ramstead, 2020). Moreover, the mismatch between the latent cause and the data is accompanied with a measure of precision. Precision is determined by the strength of the mapping between the hidden state - whose probability we are trying to make an inference about - and the incoming data (Parr et al., 2018). For example, if the mapping is accompanied with uncertainties then incoming observations have less influence on one's beliefs. Another belief characteristic closely related with precision is confidence. In Bayesian approach, confidence is altered by using incoming

9

observations to update prior beliefs - making the generative model and choices as optimal as possible – which leads to a decrease of free energy. Meyniel, Sigman & Mainen (2015) proposed that, from the Bayesian viewpoint, confidence is based on the compatibility between prior belief and observations. Authors added, that it might also be seen as a degree of belief reliability. Yanagisawa, Kawamata & Ueda (2019) went a step further by showing time-dependent subjective uncertainty (i.e. confidence), precision of belief and prediction error also interact with evidence accumulation. They showed that during earlier phases of learning, the smaller the precision of belief (i.e. greater uncertainty) the greater the evidence gain, even when the prediction error is small or close to zero. In contrast, in conditions where the prediction error is large, as individuals gain more evidence it results in an uncertainty decrease. It could be said that confidence gives value to stored beliefs, as more value will be prescribed to beliefs generating choices and actions we are more confident about. Mentioned finding is related with the assumption that confidence and uncertainty are inverse to one another, meaning that as confidence increases – uncertainty decreases. This outcome can be seen as a positive one, as reduction of uncertainty is related to positive affective states (Meyniel et al., 2015).

Figure 2. Transition from one state to the other can be influenced by the agent’s behavior (π). In the initial stage (i.e. a prior) beliefs about courses of action will depend on a prior (E), expected free energy (G), which can be divided in risk (C) and ambiguity. Another parameter that can be added here is the precision (γ) of the expected free energy.

10

Social interactions are noisy and filled with uncertainty, such that tracking one's confidence („How sure or unsure are you?“) becomes important for interacting with your environment, instead of exploiting only the „best“ option (Brown & Brune, 2012). However, before explaining the hierarchical generative model, we are going to introduce two components of trait inferences: priors, which individuals use to predict which trait is the most likely cause of what we observe (based on our prior experience) and observable evidence an agent gathers during social interactions.

2.2. Selecting socially relevant evidence

In order to successfully function in the social world, we need to know which stimuli are socially relevant, and which ones we can use as a valid source, enabling us to make inferences about hidden states (i.e. personality traits; Brown & Brüne, 2012). Moutoussis et al. (2014) hypothesized that individuals look for significant observations in order to update their beliefs about other individuals, by directing their attention to pre-selected sources (Atkinson, Simpson & Cole, 2017).

According to Connors & Halligan (2015), beliefs about other people's traits can be based on data coming from different origins. They do not have to be based on directly observable evidence, as it has been shown that information received from other people can also influence belief formation about another individual (King-Casas, Tomlin, Anen, Camerer, Quartz, & Montague, 2005). Howes (2011) added to it by pointing out that over our lifecycle we mostly learn in a social context (e.g. from other people), and not just by using affordances. In conclusion, Frith & Frith (2006) propose there are three types of information individuals use when learning about other people as individuals: direct interaction, observing other people and how they interact with their environment, and information we receive from other people. During direct interaction and by observing other individuals, there are more outcomes we can use as significant sources of evidence. For example, it was shown that individuals rely on different types of evidence (i.e. actions, underlying intentions, presence of emotional responses and facial expressions), which involve different complexity levels of cognition (i.e. perception, the theory of mind, and others) (Kestemont, Ma, Baetens, Van Overwalle & Vandekerckhove, 2015). Once individuals aggregate relevant evidence about the other person, they average over the behavioral variation due to different situations. After the average is made, they extract it as the personality of the other person – making a higher-order inference (Park, De Kay & Kraus,

11

1994). However, agents do not value all social evidence in the same manner. Based on the origin (i.e. observing other people, direct interaction or through an informant) of evidence about other people, authors have shown that certain characteristics of the context and evidence can moderate the weight individuals prescribe to observable evidence. For example, authors have shown that individuals weigh information differently based on their valence (Hughes, Zaki, & Ambady, 2017), level of compatibility with previously held beliefs (Powers, Mathys & Corlett, 2017), quality of evidence and the learning environment (Nassar, Wilson, Heasly & Gold, 2010). Meaning, the value of observable evidence is dependent on the context in which they are presented. However, the question that remains is – how do individuals know which evidence they need to attend to? Wagner, Haxby & Heatherton (2012) were also interested in this phenomenon and proposed that individuals attend only to evidence that exemplify pre-defined traits. Authors further proposed this attentional drift might be controlled by priors about trait-behavior associations.

2.3. Priors involved in social inference

Priors can be relatively open to revision and are based on previous experience and knowledge we have integrated into our memory. However, they can also be based on deeply embedded preferences or model structures that are hard-wired through evolution and our upbringing (Otten, Seth & Pinto, 2017). It can be said that priors provide a lens through which we interpret our observations of others. This enables individuals to go beyond the evidence given, making inferences about unobservable states (e.g., traits; Bruner, 1957).

The human ability to make inferences about higher-order constructs is widely affirmed in social psychology. For example, Vonk (1994) proposed that individuals implicitly identify an underlying trait when exposed to a certain behavior, suggesting they use conceptual models that explain how latent (i.e. internal) states relate to observable behavior (trait-behavior associative model). Accordingly, they can use that knowledge to infer internal states or predict future actions of another person, by using the model as input when calculating prior distributions (Bach & Schenke, 2017). According to previous literature, trait-behavior associative models or general personality models are a result of updating through experience and have a top-down effect during social interaction (Hassabis et al., 2013). More specifically, it can be said that priors, in the form of trait-behavior associations, provide an attentional drift to specific

12

observable evidence, which facilitate inferences about higher-level mental states (i.e., traits or intentions).

While prior beliefs can be hard-wired through our upbringing, we can also update our priors iteratively as we encounter new evidence. Yesterday's posterior beliefs can become today's priors, which are again subject to belief-updating. This iterative process can go indefinitely, or until a certain threshold is reached, meaning that after a certain period, individuals settle for one belief over the other. In conclusion, posterior beliefs can influence the priors for future interactions. Mentioned findings are confirmed by other authors (Asch, 1946; Wyer, Lambert, Budesheim & Gruenfeld, 1992), where it was found that already observed behavior can affect the way new evidence is used for belief updating. Here, we talked about person-based inferences, however, individuals can also use category-based evidence as priors (Chaiken & Trope, 1999), which are interrelated with person-based inferences. Categorical prior knowledge about individual groups (based on sex, race, age, etc.) can also serve as top-down filters affecting model parameters and posterior beliefs. However, we will go more into them in the following sections (for more see Chaiken & Trope, 1999; Park, 1986).

Based on these findings, we can conclude that social cognition is influenced by prior beliefs on different levels. On first level, there are general beliefs we have developed during our upbringing, such as trait-behavior associative networks, cultural and social prior knowledge (i.e. cultural/social norms and social desirability). Secondly, agents can us posterior beliefs as new priors in future interaction. However, even when priors and observable evidence are considered, there is still one component left to include – biases.

2.4. Biases involved in belief updating process

Although individuals tend to accumulate information in a nearly optimal Bayesian way (Daunizeau, Den Ouden, Pessiglione, Kiebel, Stephan, & Friston, 2010; Corradi-Dell'Acqua et al., 2015), any degree of bias can be encoded in their generative models. In this review, we will define biases simply as an a priori confidence in one belief or one mode of reasoning over another.

One of the best-known social bias is the fundamental attribution error, which suggests that people give more importance to personality traits when explaining behaviors of others and tend to neglect the context of the behavior (Ross, 1977). Individuals are more prone to make this

13

kind of error when exposed to confirmatory evidence (Park, 1989), leading to a faulty prediction of other's behavior in future settings. For example, if you make an observation „Marta did not come to my graduation“, we can use that observation to infer that „Marta is not

a reliable friend“, even though the context of the situation was that Marta did not come to our

graduation because she was working that day. Consequently, in future situations this belief will affect our inferences, making us more attentive to observations that suggest Marta is unreliable. Mentioned example leads us to another bias which is related to motivated reasoning - confirmation bias (Wason, 1960). This bias incorporates the tendency to process, search for and interpret information in a way that goes along with our prior beliefs. Schenke et al. (2016) showed individuals tend to produce confirmation bias when observing people's behavior. More specifically, they found that participants tend to act according to previously held beliefs. Other literature added to these findings (Bach & Schenke, 2017; Hughes et al., 2016), proposing it is not wrong to say that confirmation bias can be incorporated in the generative model. A third type of bias, closely associated with confirmation bias, relates to motivated reasoning due to our preferences and desired states (Teng & Kravitz, 2019). For example, Gesiarz, Cahill & Sharot (2019) showed that our preferences can affect evidence accumulation. Authors showed that when participants were encountered with desirable evidence, they needed less interactions for updating their initial beliefs, in comparison to undesirable evidence. Fourth bias can also contribute to alterations in outcome probability calculations - halo effect. Halo effect represents a bias in information processing where we generalize a belief in one domain to another domain. One of the most famous examples of the halo effect is the what-is-beautiful-is-good experiment (see Goldman & Lewis, 1977). Other types of biases, such as anchoring, representativeness, base rate neglect can lead to deviations from near optimal Bayesian inference (see Kahneman, 2013). Mathematically, these biases can be incorporated by including inverse temperature parameters in Bayesian inference, where each bias can be seen as an exponential bias for a corresponding distribution – projected as biases for prior probability and likelihood (Matsumori, Koike & Matsumoto, 2018).

Being aware of this kind of biases in belief updating is important, as they are mostly automatically integrated during belief updating. Meaning, individuals are usually not aware of their underlying influence. These types of categorizations and simplifications can be evolutionary efficient as they demand less energy consumption, making signal-transferring quicker and easier. However, it can also bring to the negative side - stereotyping and prejudice

14

(Quinn & Rosenthal, 2012). However, the question is how all of these mentioned inputs come together, yielding concise internal models.

2.5. From traits to behavior to internal models

During social interactions, individuals tend to use observable evidence to confirm or update their priors and decrease the uncertainty about the other person (Otten et al., 2016). Meaning that the goal of every individual is to decrease the amount of uncertainty about the other person by maximizing the number of evidence or optimizing the generative model. It is acknowledged that individuals can make categorizations of person-specific knowledge, considering identity, sex, race, religion, posture, and traits, which makes social learning and inferences noisy (Quinn & Rosenthal, 2012). However, in this review, we are mostly orienting on answering questions on trait-behavior inferences and seeing how they give rise to higher-level identity constructs. Here, we propose that individuals make observations to update their beliefs about other people’s traits, leading to a more concise person-specific construal (i.e. person-specific beliefs related with identity). In return, this construal will help them predict other people's behavior in the long-term, resulting in less uncertainty. Even though, they did not go into investigating the evolution of inferences about other people, Hasabbis and colleagues (2013) acknowledged that individuals can make trait inferences (i.e. extraversion or agreeableness) from behavioral vignettes. Bach & Tipper (2007) added to these findings by showing individuals create implicit personality judgments when presented to compatible actions. For example, when presented with an image of other people typing on a computer or playing football, they were more willing to judge them as „academic“ or „sporty“. Chang and colleagues (2010) went a step further by showing that observed evidence and initial other person evidence dynamically affect trait inferences – leading to an update after each new observation.

Making inferences about unobservable states (i.e. traits) from associated observable variables (i.e. actions) has certain benefits. Firstly, it enables us to predict one’s actions and behavior, likely resulting in increased evolutionary fitness. Evolutionary fitness can be defined as the likelihood of survival and reproduction in each environment (Buss, 2015). For example, being able to predict whether an individual will be loyal or betray us had an important value throughout the years (Buss, 2015). Secondly, it decreases the level of uncertainty we experience in our social environment daily. However, the timescale of reaching a certain precision of belief

15

can vary greatly, especially in noisy social environments. Many researchers tried to investigate the evolution of belief about other people's traits. For example, by using a trust game, King-Cases and colleagues (2005) found that one of the strongest predictors of trustworthiness was the amount of money players sent back to their partners. Trust game is a paradigm based on the economic theory, where there is an „Investor“ with a certain „wage“. He can invest a certain wage W1 to the „Trustee“. Now, the „Trustee“ can choose how much he will keep or return (W2) to the Investor. Trial-by-trial iterations allow the players to update his prior belief about the likelihood of two alternative outcomes, cooperate (W2 > W1) or bail (W2 < W1), and consequently result in an inference about the level of trustworthiness. In other words, we can infer about trustworthiness of others based on their reciprocal behavior. In the optimal situation, at the beginning of the game our prior likelihood is ideally equal for both options (i.e. uniform distribution) (see figure 1), as we do not have a lot of information. However, Chang and colleagues (2010) showed that we can manipulate the priors, by giving relevant information about the other individuals and this manipulation can affect predictions of the first move in the Trust game.

Mentioned findings confirm the assumption that in repeated interactions, individuals can make trait inferences based on the history of another person's behavior. However, the open question is how this history is stored. Going in the same direction, Kumaran, Summerfield, Hassabis & Maguire (2009) proposed individuals need to have a system which enables the integration of all the commonalities from gathered evidence, creating an individualized network of beliefs. Ultimately, agents will use this network to base their person- specific model on. This internal model can be seen as a formal belief space where individuals store all of their beliefs about the other person (i.e. sex, traits, gender, preferences), where every belief about a certain characteristic is themselves (implicitly) represented. In order to retain a specific belief, every belief needs to reach a certain level of precision and be held true by the individual. Similar hypotheses were suggested and confirmed by other authors (see Hassabis et al., 2013; Moutoussis et al., 2014; Park, De Kay & Kraus, 1994). It is proposed that this belief space is a result of iterative social interactions and is automatically re-activated every time we meet (i.e. identify) the person (Schenke et al., 2016). Additionally, as mentioned in previous sections, this model can inform our belief updating process in the form of prior beliefs, by affecting the likelihood individuals prescribe to certain outcomes. However, literature on specific

16

relationship between these two levels: high (identity inference) level and low (active inference) level is relatively scarce.

17

3. Hierarchical person-specific generative model

As we get to know a person, our private generative model of their behavior increases in complexity. However, a proposition for how such complexity emerges is still missing in the literature. More specifically, most literature on person-specific knowledge propose that when learning about other people we tend to store our person-specific belief in memory as a person construal. However, the question is how are these construals incorporated during time- and context-specific active inference.

It is known that in initial stages, individuals are mostly relying on incoming evidence and prior beliefs. Meaning, the central identity construct is still relatively flexible and at the beginning of its formation (Vonk, 1994; Wyer et al. 1992). Isomura, Parr & Friston (2019) also showed that prior to learning ability to categorize incoming evidence based on identity is disabled, as agents still haven't learned person-specific parameters that would help them in this categorization. However, after a couple of encounters individuals tend to notice regularities in other's behavior, enabling them to make better predictions of future outcomes, as well as a better individualized calibrations of model parameters. Meaning, context-general active inference evolves into a more individualized generative model allowing for context-specific active inference, which is dependent on identity perception. Even though there is a possibility that people use these complex models every time they meet the other person, in this literature review, an alternative solution of two-level belief update will be proposed, by invoking the hierarchical generative model introduced by Hesp and colleagues (2020). By using the hierarchical structure, we can extract the person-specificity and translate it to a higher-level, breaking it into two: context-specific active inference (lower-level) and identity perception (higher-level). This separation can simplify the belief updating process, enabling individuals to generalize context-specific (lower-level) inference on any social interaction, while the identity states will provide input as empirical priors over variables (i.e. initial states, priors over precision and policies, likelihood matrices) at the lower level (Hesp et al., 2020). Empirical priors incorporate individuals' (implicit) knowledge about their dynamics with the world (Moutoussis et al., 2014), enabling the interaction between higher-level and lower-level beliefs. One of the benefits of the hierarchical model is that it enables us to combine previous research findings about person-specific construals with the predictive brain hypothesis. Furthermore, hierarchical structure allows individuals to accumulate evidence about other people, store occurrences over multiple interactions and update person-specific parameters between levels in

18

a temporally nested way - state transitions on the higher-level happen across interactions and lower-level state transitions happen within interactions. In our case, both levels are co-dependently involved in belief updating by exchanging top-down and bottom-up information.

However, the question that remains is how do we “perceive” someone’s identity and how does a person know which model to update? Isomura and colleagues (2019) suggested a solution to these questions. Authors proposed a scheme which enables fitting sensory input under each generative model by inferring which generative model is related with sensory input (i.e. facial features) and defining the model plausibility parameter. Model plausibility parameter can be thought of as a switching parameter, that defines the likelihood that the specific model is generating the input. Mentioned parameter is also important in learning, as it weights on the learning rate, leading to updates of parameters of most plausible models. In our case, each person-specific generative model can be seen as a hypothesis about someone’s identity, which will be affected by the dynamics of the lower-level, as this level provides bottom-up evidence (i.e. facial features), allowing the agent to update the model plausibility parameter (i.e. recognize someone or “perceive” their identity). Once the most plausible person-specific generative model is defined, it will provide top-down priors for context-specific active future inferences.

Even though this alternative solution is still an exploratory hypothesis, it can be used to account for previously mentioned social phenomena. For example, it can help in explaining how individuals are transitioning from using only prior person-specific beliefs and incoming evidence to relying on higher-order person-specific beliefs evoked by identity inferences. Furthermore, it can account for peoples’ ability to switch between different models when learning about other peoples’ behavior (i.e. update parameters of person-specific generative models). The form of this hierarchical model already derives support from other research on time-dependent evidence accumulation (Bach & Schenke, 2017; Schenke et al., 2016; Park, 1986) and integration of conflicting information (Hughes et al., 2016). Moreover, Hassabis and colleagues (2013) argued that having a higher level that provides prior knowledge about the other person (i.e., an identity-personality model) is essential for predicting and understanding their behavior. However, more research in this field is needed.

19

4. Person-specific priors and expectations modulate our actions and

reality

So far, we have reviewed many variables involved in the formation of person-specific beliefs, from our prior socio-cultural knowledge, preferences, and biases to our direct observations (King-Cases et al. 2005). Meaning, all things constant (i.e. situation, incoming evidence) individuals can end up with different beliefs and expectations (Fryer, Harms & Jackson, 2017). However, even though our posterior beliefs will always be influenced by our preexisting knowledge, preferences, goals and cultural priors – their influence decrease as the number of interactions increase (Fryer et al., 2017).

Expectations about outcomes in social situations can be based on person-specific beliefs. Such expectations have a high evolutionary benefit, as they increase our grip on our dynamic social environment. More specifically, iterative nature of belief updating during social interactions enables individuals to reduce their uncertainties about the other person, as much as it leads to a more optimized generative model. However, successive interactions can also cause robusness of person-specific beliefs, leading to a default and automatic state of mind – „status quo“. The problem appears if individuals start to generalize this state of mind and use it as a starting point in other social interactions - not being able to critically evaluate incoming evidence. Additionally, people strive for consistency, meaning they are more motivated to give priority to incoming evidence that are in line with existing beliefs – making the belief even more fixed. So, once we think we „know“ the person, expectations are made implicitly. For example, by knowing a person for a very long time, you hold very precise person-specific beliefs, which are also related with your expectations about the outcome and, consequently, your behavior. Imagine that you hold a very strong belief about your friend being extroverted, rather than introverted. This belief can map onto your expectations (i.e. They will go on a party with us), which will consequently affect your behavior – showing approaching behavior by inviting them to the party. However, if we believe they are introverted, our expectations will change accordingly, resulting in avoiding behavior (i.e. not inviting her to the party). The outcome we observe will have an impact on the precision we prescribed to that belief. Meaning, if the outcome is not in accordance with our prior belief, we need to revise it. However, the impact of the prediction error on belief revision is time dependent - as time and number of interactions increase, person-specific beliefs become more robust to contradictory observations

20

(i.e. not leading to belief or behavior updating). From this example, we can see how our prior beliefs and expectations can affect our behavior and consequently our reality. Also, there are findings (Puviani & Remi, 2016; Yanagisawa et al., 2019; Hesp et al., 2020) that expectations can regulate our emotions as well. For example, Yanagisawa et al., (2019) showed that the effect of prediction error on emotions also depends on the stability of our expectations. More specifically, the more certain our expectations the higher the effect of prediction error on emotional reactions (i.e. surprise or disappointment). From these finding, we can conclude there is an indirect relationship between our expectations and affective responding, where significant mismatches in expectations (i.e. prior beliefs) and novel evidence have the power to generate specific emotions (Hoogendoorn, Treur, van der Wal & van Wissen, 2010). Even though these findings were presented on an individual level, the mentioned analogy can also be translated to inferences on a group level, by changing person-specific identity to group-specific identity. However, research investigating the role of expectations in social interactions is still very scarce, leading to a need for more research investigating the effect of expectations on emotional reactivity and action towards different groups of people.

One implication worth investigating relates to the effect of social distance between people (i.e. strangers, friends and family). Social distance can be defined as an affective closeness between us and other people (Trope & Liberman, 2010). Investigating the role of social distance is not novel, as there is some research on the effect of social distance on decision making and emotional arousal (Sun, Liu, Zhang & Lu, 2016). For example, Jung, Sul & Kim (2013)showed that brain regions related to emotion processing are more activated when making decisions for individuals socially closer to us, in contrast to socially distant individuals (Jung et al. 2013). Moreover, Mandel (2006) showed that individuals have different justice expectations from friends, in comparison to strangers. Additionally, the author found expectations determined their willingness to agree on their offers, where they were more willing to accept the offer from a friend, in contrast to strangers. Mentioned findings might suggest that we integrate information in a different way when social distance is considered. However, a second possible explanation is that with friends and family we have more interactions than with someone we just met. These repetitive interactions enable us to gather more evidence – leading to higher precision, more confidence and robustness of the person-specific beliefs. Consequently, higher confidence and robustness of beliefs will result in a switch, where individual's behavior will be more influenced by their expectations (i.e. prior beliefs), rather than incoming information.

21

More specifically, with friends and family individuals are more susceptible to illusions of knowledge, making them think they can correctly predict their actions. On the other hand, when meeting a stranger, the uncertainty of beliefs is still very high, which means, the optimal way of getting to know them is to gather as much incoming evidence as possible, in contrast to only relying on their prior beliefs. Third explanation can be that beliefs and expectations about friends and family members start to interact with group-specific beliefs (i.e. family and friend). Meaning, our expectations can also be affected by cultural and social norms of what is expected from people that we label as family and friends. For example, have you ever found yourself confident your friend or family member will do something you expect, just because they are your friend of family member? Mentioned explanations give us an idea of how we can use previously mentioned knowledge on person-specific beliefs as a guide for future research.

However, this knowledge can be expanded to other phenomena we encounter in our social environment. For example, self-fulfilling prophecy occurs when certain social belief results in the individual acting in ways which will confirm that belief (Jussim, 2001). Importantly, Biggs (2009) noted that individuals, within the process, are not aware of how their beliefs help in constructing their reality. Having that in mind, another plausible implication of expectation research is expectation management. In organizational psychology, expectation management is defined as „...management of knowledge extracted from expectations from another individuals, as they resonate from their experience, information, ideas and personal interest.“ (Luomo-aho & Olkkonen, 2016). Based on everything mentioned, it is visible there is a certain amount of noise in belief updating that give rise to irrational expectations, that can be managed. Expectation management can be used in a broader sense in relational therapy, cognitive-behavior therapy, social relations & organizations. Moreover, by moderating expectations (i.e. reference points) and keeping them closer to reality (i.e. objective information) we are able to controle the size of prediction errors, leading to less negative and more positive surprises. Even though, there are a lot of possible applications of expectation management, it is still in its beginnings and has a long way to be explored in full.

In this section we wanted to depict the power of expectations and to address how this knowledge can be applied in a practical way. As there is still a substantial lack of research on the way our expectations impact our cognition on a higher level, we are only able to make assumptions about the details of their interaction. However, another field we can use to help us in this area are findings within the neuroscience framework.

22

5. Neural correlates of Bayesian active inference in social cognition

Looking from a neuroscientific point of view, studies have shown there is an overlap among brain areas involved in error monitoring, encoding PE and social decision-making (see Brown & Brune, 2012). Mentioned finding highlights the importance of predicting mechanisms in social decision making.

When it comes to explaining the Bayesian predictive brain hypothesis, most research focus on lower-level cognition, such as visual perception (Otten et al., 2016). Yet, relatively little is known about which brain regions are involved in active inference (i.e. interaction of priors and external information) during higher-order cognition. Nevertheless, the universal idea is that neural signals reflect the mismatch between expected beliefs/values and incoming evidence (i.e. prediction error) (Parr et al., 2018). In our case, predictor neurons - neurons that format predictions for incoming evidence - encode expectations about the incoming evidence, which they forward to dopaminergic neurons (i.e. error neurons). In situations where incoming evidence matches the person-specific prior expectation, encoded in prediction neurons, the response of error neurons is reduced. On the other hand, when the incoming data deviates from the expected, the response of error neurons is enhanced (Koster-Hale & Saxe, 2013). Similar activation patterns are also recorded in the VS, which is associated with belief precision (Matamales et al., 2019; Friston et al., 2017; Hesp et al., 2020). Mentioned findings demonstrate that the evaluation of incoming evidence is relative, as it is always compared with out priors and expectations we have about other people. Connors & Halligan (2015) added to this by suggesting that once we accept person-specific beliefs and hold them as true, they can have an overall impact on the cognitive system, contributing and affecting our perception, memory and behavior.

Based on previously mentioned literature, we can see that social learning requires conscious retention of occurrences and posterior person-specific beliefs, based on which incoming evidence can be compared. Kumaran and colleagues (2009) proposed that this retention of occurrences is done by finding commonalities among related experiences, which result in associational knowledge networks. Furthermore, authors found that the connectivity between the hippocampus (HC) and the vmPFC might be involved in constructing associations between posteriors and novel evidence. In this connectivity, HC has an important role in choice making by forwarding memory signals of possible options to prefrontal regions (i.e. vmPFC and

23

orbitofrontal cortex) for evaluation. These findings indicate that efficient connectivity between specific parts of the brain is more important than looking at brain regions in isolation (see Wang & Morris, 2009; Wybo, Torben-Nielsen, Nevian & Gewaltig, 2019). Hassabis and colleagues (2013) also suggested a role of the mPFC in belief encoding. Additionally, they found different activation patterns in the anterior mPFC when subjects made trait inferences about different people, suggesting that people have an ability to differentiate between people’s identities. On a lower level, Stern, Gonzales, Welsh & Taylor (2010) introduced two additional brain regions that might be involved in person-specific belief updating – dorsal anterior cingulate cortex (dACC), which is also important for error monitoring (Matsumoto & Tanaka, 2004) and ventromedial prefrontal cortex (vmPFC). However, the question is how these beliefs are stored in the brain. Wang & Morris (2010) suggested it might be seen as frameworks of knowledge represented as an interconnected network of cortical representations, which is carried out by synaptic plasticity and the connectivity between previously mentioned brain regions. In return, this network can help in expectation generation, which individuals use when predicting outcomes for future interactions. Presented finding was also supported with newer research on memory and neural signal processing (see Vaz,Witting, Inati & Zaghloul, 2020; Wybo et al., 2019).

Lastly, mentioned assumption about the relationship between PE and emotional reactivity also found a confirmation on a neural level. For example, Puviani & Rama (2016) proposed orbitofrontal cortex (OFC) encodes the expected outcome and evaluates the prediction error generated within the ventral striatum (VS), leading to a build-up of an emotional response in the amygdala. As a result, amygdala causes a neural emotional response determined by the amount of the discrepancy between the expected outcome and the objective outcome. Based on these findings we can see that there is a suggestive relation between the prediction error and emotional reactivity.

In conclusion, by adding the overview of research in neuroscience, we can see how these findings can be used as a validation of previously mentioned behavioral findings. Furthermore, it enables us to pin-point certain areas that are active during Bayesian social inferences and can be used as a support for the suggested hierarchical inference structure. However, more studies combining social neuroscience, computer science and social psychology are needed.

24

6. Conclusion

The main aim of this review was to present the „lifecycle“ of person-related beliefs and how they come together to affect our behavior and reality. Moreover, we wanted to use this literature review as a brief introduction of how we can use the combination of computer science, social neuroscience and social psychology to represent regularities in social learning and introduce an alternative solution to combining higher-level identity inferences with lower-lever inferences, but also to depict the subjective nature of our beliefs. Hopefully, we successfully portrayed social interactions as learning environments in which individuals try to gather as much information as possible in order to reduce uncertainty about other people. Reduction of uncertainty is accomplished with successive belief updating, which is executed by combining incoming evidence and prior beliefs/knowledge following the Bayes’ rule. However, it was demonstrated that this process can be noisy, as there are biases related with motivational reasoning, preferences, experience and prior knowledge, individuals gather during their upbringing. This subjectivity becomes dangerous when beliefs reach a specific threshold of precision and become more robust and automatic. Hopefully, we managed to reflect the importance of having more research about person-specific belief formation and role of expectations, as it gives valuable insight which can be positively applied to other fields, such as organizational psychology, psychotherapy, etc.

Because of the limited space and the length of the review, we only encompassed person-specific beliefs. However, this analogy can also be translated to group-person-specific beliefs, as they can also have an impact on formatting impressions about individuals (Quinn & Rosenthal, 2012). Additionally, we did not put much emphasis on theory of mind which can also be added to the equation of social inferences. Furthermore, all the examples followed Bayesian framework. However, they are not based on precise mathematical analogies, as they were used more as a way for getting across all the literature. For example, we did not consider the difference in weights individuals prescribe to certain type of information. Nonetheless, this can be an implication for further research topics, such as evidence selection during social learning or trying to determine the way in which individuals integrate all of this information, by using novel approaches adopted from computer science. Meaning, every example needs to be considered in a critical way by using pre-existing knowledge from psychology, neuroscience and sociology.

25

7. Literature

Asch, S. E. (1946). Forming impressions of personality. The Journal of Abnormal and Social Psychology, 41(3), 258.

Ambady, N., Bernieri, F. J., & Richeson, J. A. (2000). Toward a histology of social behavior: Judgmental accuracy from thin slices of the behavioral stream. In Advances in experimental social psychology (Vol. 32, pp. 201-271). Academic Press.

Atkinson, M. A., Simpson, A. A., & Cole, G. G. (2018). Visual attention and action: How cueing, direct mapping, and social interactions drive orienting. Psychonomic bulletin & review, 25(5), 1585-1605.

Bach, P., & Schenke, K. C. (2017). Predictive social perception: Towards a unifying framework from action observation to person knowledge. Social and Personality Psychology Compass, 11(7), e12312.

Bach, P., & Tipper, S. P. (2007). Implicit action encoding influences personal-trait judgments. Cognition, 102(2), 151-178.

Baker, C., Saxe, R., & Tenenbaum, J. (2011). Bayesian theory of mind: Modeling joint belief-desire attribution. In Proceedings of the annual meeting of the cognitive science society (Vol. 33, No. 33).

Bolis, D., & Schilbach, L. (2017). Beyond one Bayesian brain: Modeling intra-and inter-personal processes during social interaction: Commentary on “mentalizing homeostasis: The social origins of interoceptive inference” by Fotopoulou & Tsakiris. Neuropsychoanalysis. Taylor and Francis Ltd. https://doi.org/10.1080/15294145.2017.1295215

Brod, G., Werkle-Bergner, M., & Shing, Y. L. (2013). The influence of prior knowledge on memory: a developmental cognitive neuroscience perspective. Frontiers in behavioral neuroscience, 7, 139.

Brown, E. C., & Brüne, M. (2012). The role of prediction in social neuroscience. Frontiers in human neuroscience, 6, 147.

26

Bruner, J.S. (1957). Going beyond the information given. In J.S. Bruner, E, Brunswik, L. Festinger, F. Heider, K.F. Muenzinger, C.E. Osgood, & D. Rapaport, (Eds.), Contemporary approaches to cognition(pp. 41-69). Cambridge, MA: Harvard University Press.

Buss, D. (2015). Evolutionary psychology: The new science of the mind. Psychology Press. Chaiken, S., & Trope, Y. (Eds.). (1999). Dual-process theories in social psychology. Guilford Press.

Chang, L. J., Doll, B. B., van’t Wout, M., Frank, M. J., & Sanfey, A. G. (2010). Seeing is believing: Trustworthiness as a dynamic belief. Cognitive psychology, 61(2), 87-105.

Chang, L. J., & Koban, L. (2013). Modeling emotion and learning of norms in social interactions. Journal of Neuroscience, 33(18), 7615-7617.

Christian, B., & Griffiths, T. (2016). Algorithms to live by: The computer science of human decisions. Macmillan.

Connors, M. H., & Halligan, P. W. (2015). A cognitive account of belief: A tentative road map. Frontiers in psychology, 5, 1588.

Corradi-Dell'Acqua, C., Turri, F., Kaufmann, L., Clément, F., & Schwartz, S. (2015). How the brain predicts people's behavior in relation to rules and desires. Evidence of a medio-prefrontal dissociation. cortex, 70, 21-34.

Daunizeau, J., Den Ouden, H. E., Pessiglione, M., Kiebel, S. J., Stephan, K. E., & Friston, K. J. (2010). Observing the observer (I): meta-bayesian models of learning and decision-making. PloS one, 5(12).

De Martino, B., Fleming, S. M., Garrett, N., & Dolan, R. J. (2013). Confidence in value-based choice. Nature neuroscience, 16(1), 105.

El-Deredy, W., Trujillo- Barreto, N. J., Watson, A., & Jones, A. K. (2010). Bayesian model comparison of placebo analgesia: Parameterisation of priorinformationand expectation.IASP 2010. Montreal.

Frith, C. D., & Frith, U. (2006). How we predict what other people are going to do. Brain research, 1079(1), 36-46.

27

Friston, K. J., & Stephan, K. E. (2007). Free-energy and the brain. Synthese, 159(3), 417-458.

Fryer Jr, R. G., Harms, P., & Jackson, M. O. (2019). Updating beliefs when evidence is open to interpretation: Implications for bias and polarization. Journal of the European Economic Association, 17(5), 1470-1501.

Gesiarz, F., Cahill, D., & Sharot, T. (2019). Evidence accumulation is biased by motivation: A computational account. PLoS computational biology, 15(6).

Gilboa, A., Alain, C., Stuss, D. T., Melo, B., Miller, S., & Moscovitch, M. (2006). Mechanisms of spontaneous confabulations: a strategic retrieval account. Brain, 129(6), 1399-1414.

Goldman, W., & Lewis, P. (1977). Beautiful is good: Evidence that the physically attractive are more socially skillful. Journal of Experimental Social Psychology, 13(2), 125-130.

Hassabis, D., Spreng, R. N., Rusu, A. A., Robbins, C. A., Mar, R. A., & Schacter, D. L. (2013). Imagine all the people: how the brain creates and uses personality models to predict behavior. Cerebral Cortex, 24(8), 1979-1987.

Hauser, T. U., Eldar, E., & Dolan, R. J. (2017). Separate mesocortical and mesolimbic pathways encode effort and reward learning signals. Proceedings of the National Academy of Sciences, 114(35), E7395E7404.

Heleven, E., & Van Overwalle, F. (2015). The person within: memory codes for persons and traits using fMRI repetition suppression. Social Cognitive and Affective Neuroscience, 11(1), 159-171.

Hesp, C., Smith, R., Allen, M., Friston, K., & Ramstead, M. (2020). Deeply felt affect: the emergence of valence in deep active inference.

Hoogendoorn, M., Treur, J., Van Der Wal, C. N., & Van Wissen, A. (2010). Modelling the interplay of emotions, beliefs and intentions within collective decision making based on insights from social neuroscience. In International Conference on Neural Information Processing (pp. 196-206). Springer, Berlin, Heidelberg.

28

Hughes, B. L., Zaki, J., & Ambady, N. (2017). Motivation alters impression formation and related neural systems. Social cognitive and affective neuroscience, 12(1), 49-60.

Isomura, Takuya, Thomas Parr, and Karl Friston. "Bayesian Filtering with Multiple Internal Models: Toward a Theory of Social Intelligence." Neural Computation 31, no. 12 (2019): 2390-2431.

Jung, D., Sul, S., & Kim, H. (2013). Dissociable neural processes underlying risky decisions for self versus other. Frontiers in neuroscience, 7, 15.

Kahneman, D. (2013). Thinking, fast and slow. London: Penguin Group.

Kestemont, J., Ma, N., Baetens, K., Clément, N., Van Overwalle, F., & Vandekerckhove, M. (2015). Neural correlates of attributing causes to the self, another person and the situation. Social cognitive and affective neuroscience, 10(1), 114-121.

Klein S.B. & Loftus J. (1990). The role of abstract and exemplar-based knowledge in self-judgments: Implications for a cognitive model of the self. Advances in Social Cognition., 3:131–139.

King-Casas, B., Tomlin, D., Anen, C., Camerer, C. F., Quartz, S. R., & Montague, P. R. (2005). Getting to know you: reputation and trust in a two-person economic exchange. Science, 308(5718), 78-83.

King-Casas, B., Sharp, C., Lomax-Bream, L., Lohrenz, T., Fonagy, P., & Montague, P. R. (2008). The rupture and repair of cooperation in borderline personality disorder. science, 321(5890), 806-810.

Koster-Hale, J., & Saxe, R. (2013). Theory of mind: a neural prediction problem. Neuron, 79(5), 836-848.

Kumaran, D., Summerfield, J. J., Hassabis, D., & Maguire, E. A. (2009). Tracking the emergence of conceptual knowledge during human decision making. Neuron, 63(6), 889-901.

LaBerge, D. (1995). Perspectives in cognitive neuroscience. Attentional processing: The brain's art of mindfulness. Harvard University Press.

29

Lebreton, M., Abitbol, R., Daunizeau, J., & Pessiglione, M. (2015). Automatic integration of confidence in the brain valuation signal. Nature neuroscience, 18(8), 1159.

Lind, J., Ghirlanda, S., & Enquist, M. (2019). Social learning through associative processes: A computational theory. Royal Society open science, 6(3), 181777.

Luoma-aho, Vilma & Olkkonen, Laura. (2016). Expectation management. In C. E. Carroll (Ed.) The SAGE Encyclopedia of Corporate Reputation (pp. I:303-306). Thousand Oaks, CA: SAGE.

Nassar, M. R., Wilson, R. C., Heasly, B., & Gold, J. I. (2010). An approximately Bayesian delta-rule model explains the dynamics of belief updating in a changing environment. Journal of Neuroscience, 30(37), 12366-12378.

Macrae, C. N., & Bodenhausen, G. V. (2000). Social cognition: Thinking categorically about others. Annual review of psychology, 51(1), 93-120.

Mandel, D. R. (2006). Economic transactions among friends: Asymmetric generosity but not agreement in buyers' and sellers' offers. Journal of Conflict Resolution, 50(4), 584-606.

Matamales, M., McGovern, A. E., Dai Mi, J., Mazzone, S. B., Balleine, B. W., & Bertran-Gonzalez, J. (2019). D1 and D2 systems converge in the striatum to update goal-directed learning. bioRxiv, 780346.

Mathys, C., Daunizeau, J., Friston, K. J., & Stephan, K. E. (2011). A Bayesian foundation for individual learning under uncertainty. Frontiers in human neuroscience, 5, 39.

Matsumori, K., Koike, Y., & Matsumoto, K. (2018). A biased Bayesian inference for decision-making and cognitive control. Frontiers in neuroscience, 12, 734.

Mercier, H. & Sperber, D. (2018). The enigma of reason. UK: Penguin Random House. Meyniel, F., Sigman, M., & Mainen, Z. F. (2015). Confidence as Bayesian probability: From neural origins to behavior. Neuron, 88(1), 78-92.

Molenberghs, P., Brander, C., Mattingley, J. B., and Cunnington, R. (2010). The role of the superior temporal sulcus and the mirror neuron system in imitation. Hum. Brain Mapp. 31, 1316–1326.

30

Moutoussis, M., Fearon, P., El-Deredy, W., Dolan, R. J., & Friston, K. J. (2014). Bayesian inferences about the self (and others): A review. Consciousness and cognition, 25, 67-76.

Mrug, S., & Hoza, B. (2007). Impression formation and modifiability: Testing a theoretical model. Merrill-Palmer Quarterly (1982-), 631-659.

Oldham, S., Murawski, C., Fornito, A., Youssef, G., Yücel, M., & Lorenzetti, V. (2018). The anticipation and outcome phases of reward and loss processing: A neuroimaging meta‐ analysis of the monetary incentive delay task. Human brain mapping, 39(8), 3398-3418.

Otten, M., Seth, A. K., & Pinto, Y. (2017). A social Bayesian brain: How social knowledge can shape visual perception. Brain and Cognition, 112, 69-77.

Quinn, K. A., & Rosenthal, H. E. (2012). Categorizing others and the self: How social memory structures guide social perception and behavior. Learning and Motivation, 43(4), 247-258.

Park, B. (1986). A method for studying the development of impressions of real people. Journal of Personality and Social Psychology, 51(5), 907.

Park, B., DeKay, M. L., & Kraus, S. (1994). Aggregating social behavior into person models: Perceiver-induced consistency. Journal of Personality and Social Psychology, 66(3), 437.

Parr, T., & Friston, K. J. (2019). Attention or salience?. Current opinion in psychology, 29, 1-5.

Parr, T., Rees, G., & Friston, K. J. (2018). Computational neuropsychology and Bayesian inference. Frontiers in human neuroscience, 12, 61.

Powers, A. R., Mathys, C., & Corlett, P. R. (2017). Pavlovian conditioning–induced hallucinations result from overweighting of perceptual priors. Science, 357(6351), 596-600.

Puviani, L., & Rama, S. (2016). A system computational model of implicit emotional learning. Frontiers in computational neuroscience, 10, 54.

31

Ross, L. (1977). The intuitive psychologist and his shortcomings: Distortions in the attribution process. In Advances in experimental social psychology (Vol. 10, pp. 173-220). Academic Press.

Schenke, K. C., Wyer, N. A., & Bach, P. (2016). The things you do: Internal models of others’ expected behaviour guide action observation. PloS one, 11(7), e0158910.

Schultz, W. (2007). Multiple dopamine functions at different time courses. Annu. Rev. Neurosci., 30, 259-288.

Sun, Q., Liu, Y., Zhang, H., & Lu, J. (2017). Increased social distance makes people more risk-neutral. The Journal of social psychology, 157(4), 502-512.

Stern, E. R., Gonzalez, R., Welsh, R. C., & Taylor, S. F. (2010). Updating beliefs for a decision: neural correlates of uncertainty and underconfidence. Journal of Neuroscience, 30(23), 8032-8041.

Teng, C., & Kravitz, D. J. (2019). Visual working memory directly alters perception. Nature human behaviour, 3(8), 827-836.

Ting, C. C., Yu, C. C., Maloney, L. T., & Wu, S. W. (2015). Neural mechanisms for integrating prior knowledge and likelihood in value-based probabilistic inference. Journal of Neuroscience, 35(4), 1792-1805.

Trope, Y., & Liberman, N. (2010). Construal-level theory of psychological distance. Psychological review, 117(2), 440.

Vaz, P.A., Witting, H. J., Inati, K.S. & Zaghloul, A.K. (2020). Replay of cortical spiking sequences during human memory retrieval. Science, 367(6482), 1131-1134.

Veissière, S. P., Constant, A., Ramstead, M. J., Friston, K. J., & Kirmayer, L. J. (2019). Thinking through other minds: A variational approach to cognition and culture. Behavioral and Brain Sciences, 1-97.

Vonk, R. (1994). Trait inferences, impression formation, and person memory: Strategies in processing inconsistent information about persons. European review of social psychology, 5(1), 111-149.