Author
Braunsdorf, Marius marius.braunsdorf@student.uva.nl
Universiteit van Amsterdam
1st Supervisor
Dr. Mars, Rogier r.mars@donders.ru.nl Radboud University Nijmegen
University of Oxford 2nd Supervisor Dr. Kolling, Nils nils.kolling@psy.ox.ac.uk University of Oxford UvA Representative Dr. van Elk, Michiel
m.vanelk@uva.nl November 29, 2017 Universiteit van Amsterdam
Making and inferring decisions for ourselves and others forms is a crucial part of human functioning in everyday life. Many of those decisions take place on behalf of others. The mechanisms behind those decisions, however are poorly understood. The current work gave a first overview on a computationally grounded mechanistic framework for a human Theory of Mind. While parallely accumulating evidence for oneself and other agents, participants solved a well grounded evidence accumulation paradigm in a social context. Results showed, that besides uniquely social brain regions as the tempoparietal junction or the superior temporal sulcus showed activation. Crucialll though are the findings that activity in non-social decision making regions - such as the dorsal anterior cingulate cortex, the paracingulate cortex and the medial prefrontal cortex - was scaled by the presented evidence . Those findings challenge the notion of a part of the brain being uniquely social and gather evidence in favour of the idea, that general decision making mechanisms are recruited in different (social) contexts depending on task requirements.
Keywords: Social, Decision-making, fMRI, ACC, Frontal, pre-SMA, Theory of Mind
1 Introduction 1
1.1 Theory of Mind . . . 2
1.2 Social Computation . . . 4
1.3 The current study . . . 7
2 Methods 9 2.1 Participants . . . 9 2.2 The Task . . . 10 2.3 Training . . . 12 2.4 The Experiment . . . 13 2.5 Data Collection . . . 13
2.6 Behavioral Data Analysis . . . 14
2.7 Imaging data analysis . . . 15
2.7.1 Preprocessing . . . 15
2.7.2 First-level analyses . . . 15
2.7.3 Group-level analyses . . . 18
3 Results 18 3.1 Behavioural results . . . 18
3.1.1 Regression on total evidence . . . 18
3.1.2 Regression on single sample information . . . 21
3.1.3 Regression on evidence split by agent ability . . . 23
3.1.4 Agent rating . . . 25
3.2 fMRI Results . . . 26
4 Discussion 28 4.1 Imaging results vmPFC and ACC . . . 29
1 Introduction
Among all primates humans form the biggest social groups and interact in the most complex way with other individuals of their own species (Hill & Dunbar, 2003). It can be argued that this is one of the key factors which ultimately helped us to evolve to be such a successful species, overcoming problems imposed by our environment which are not solvable individually. Socrates already recognized that “humans are by nature social animals” (Starr, 2013). Building on this millennia old statement where humans are described as ’infovores’ (not without a slight wink of the eye, of course). What is necessary to meaningfully interact with others and how do we end up with the astonishing ability to infer reasonably well what others will do, why they do it, and how they arrive at certain conclusions?
First of all one has to know that the other is a conscious being. To entertain this thought we necessarily have to infer the content of meaningful representations of others’ mental states in our own, or as it is put so pleasantly confusing in common language “I think that you think”. This idea is known as the “Theory of Mind”. From an philosophical point of view this is an indisputable necessity of human functioning. This concept is first explored in this introduction, since it lays the groundwork for the current endeavour of disentangling the mechanisms giving rise to our own knowledge and knowledge about the state of the world of others. The next sections sketches a brief history of ToM conceptualization and empirical research. A short overview of early behavioural beginnings when mind and brain were left as the famous black box are presented gradually advancing to ideas of social cognitive neuroscience. Although the concept of ToM is very broad, many studies handle a somewhat more narrow definition. The inference about the others’ state of the world does not necessarily represent the whole idea rooted in the Theory of Mind (ToM), stating that we entertain a complete representation about the others’ conscious experience. This work focusses on the behavioural predictions of the people around us, which humans infer based on the idea of similar mechanisms and thoughts giving rise to similar decisions as our own. An ability not encompassing the whole construct of ToM, but forming a basic necessity for it to emerge in the first place.
Those behavioural predictions form the cornerstone for the area of social decision making, investigating the ability to decide for oneself as well as on the behalf of others. Research done in the past as well as theoretical implications of social computation will form the second part of this introduction, before exploring the advantages and possibilities of the current study to add to the existing body of knowledge in the relatively recent field of social cognitive neuroscience.
1.1 Theory of Mind
Theory of Mind (ToM) describes the philosophical conceptualization of the human capacity to automatically engage in the idea that other people also have conscious ex-periences and cognitive capabilities such as believes, desires or knowledge, different from those of the self (Premack & Woodruff, 1978).
Early philosophical conceptualizations are rooted in the 50s of the twentieth cen-tury. They are termed the theory theory approach. This approach regards humans and especially children as scientists who try to make sense of the world around them by at-tributing certain causes to observed events. Those events can be of a broad array of different kinds, such as observable behaviour with biological causes (e.g. “Someone with a heavy physical job has to eat a lot”), but it can also include theories about believes, intentions, emotions or goals of other people (e.g. “Someone who smiles a lot is happy”). Especially the latter gives rise to the concept of a ToM in humans, where certain be-haviours are attributed to different mental states. We as humans thus infer that other people are capable of experiencing happiness. Those early conceptualizations kick-started empirical behavioural investigations about the idea in the 70s, Most noteworthy is the nowadays infamous false belief task, popularly called the “Sally-Anne task”. In this task children are introduced to two dolls, Sally and Anne. Sally then hides a marble in her box and leaves the scene. Subsequently, Anne comes along and retrieves the marble from Sally’s box and puts it in her own. The participants are then asked where Sally will most likely look for the marble. While it seems quite clear that Sally will look in her own box, this answer is not as straightforward as we might expect from a children’s perspective.
Crucially, to solve the task one has to have a representation of the mental state of the other person, which differs from the actual state of the world. While the behavioural entertainment of Sally’s state of the world can be investigated here quite perfectly, this paradigm does not afford any possibility to investigate computational mechanisms giving rise to that representation because of a missing interpretation of the processes ultimately making this representation and the difference between it and our own knowledge possible. A more recent theoretical approach paying attention to modern research methods and advances into the understanding of ToM is called the simulation theory. It proposes the idea that humans can predict the behaviour of others, by thinking about what they would do themselves in the place of that other person Gordon (1986). Goldman (1992) raised the idea, that people engage in a simulation of the others’ mental state by employ-ing the same mechanisms givemploy-ing rise to their own mental states to create additional states which are projected onto others. In this way the hypothesized mental state of another agent is modelled. While the simulation theory from a philosophical reference framework engages in high-level behaviour, such as the attribution of propositional attributes, it in-spired a considerable body of research into imitation of low-level components of cognition and behaviour.
Especially the discovery of mirror neurons in primates (Rizzolatti, Fadiga, Gallese, & Fogassi, 1996) provided the impetus for neuroscientific investigations of social behaviour and the capacity to mentalize about different aspects of another conscious being. A mir-ror neuron is active when an agent observes a certain action or carries out that action, thus activity is elicited if the behaviour is mirrored (Rizzolatti & Craighero, 2004). Mirror neuron system activity is mainly understood for simple behaviours and low-level cogni-tive phenomena such as perception or direct imitation of body movements. Theoretical frameworks developed speculating about neural resemblance of intentions rather than just imitated or inferred behavioural predictions (Goldman & Others, 2012). Evidence from fMRI experiments in humans pleading in favour of the intention based mirroring of actions is gathered by Iacoboni et al. (2005) in an experiment where participants watched hand movements of other agents grasping objects. In the two conditions (intention and
control) a significantly higher increase in activity was witnessed in pre motor areas only in the intention group compared to the control group. Since the monitoring and construc-tion of others’ believes and intenconstruc-tions is a complicated process solved with astonishing precision in humans, the current work draws on inspiration from the field of computa-tional neuroscience applied to a social paradigm. Recent advances in social neuroscience will be explained in the following section.
1.2 Social Computation
We can only have meaningful interactions with others if we grasp intentions and believes of those others implemented in the idea of a theory of mind in humans. Recent technological advances made it possible to critically assess this capacity for meaningful social interactions. In this section, empirical evidence of identified regions and mech-anisms afforded by social interaction are presented. The most fruitful approaches are discussed, closing the section with an explanation of the gaps in current knowledge about the neural cognitive basis of ToM and what the current work contributes.
It is clear now, that certain necessities have to be present for meaningful social behaviour to emerge. From a neuroscientific point of view the computational mechanisms giving rise to that behaviour form the utmost importance as a research topic. In order to carefully investigate the role of different brain regions it is necessary to decompose social behaviour into its core components and to make those core components directly quantifi-able and measurquantifi-able. In the history of computational social neuroscience a considerquantifi-able lack of careful decomposition can be witnessed and studies conducted so far were only successful in illuminating the role of distinct networks, but failed to dissect the exact roles of those networks’ parts.
Through the help of modern neuroimaging techniques two networks seem of par-ticular importance when investigating social cognition. A first network constits of the Anterior Cingulate Cortex Sulcus (ACCs) the ventromedial prefrontal cortex (VMPFC), the Amygdala and the ventral striatum (VSTr), dubbed the reward and reinforcement learning network (Behrens, Hunt, & Rushworth, 2009).
Functional imaging studies raised several hypothesis about the ACCs role in sev-eral decision making domains, with the most prominent debate being conflict resolution versus error monitoring affording activity in that region (Lütcke & Frahm, 2007). Activ-ity in the ACC however is usually widespread rendering a decomposition of functional roles within this area as highly speculative.
Dorsal parts of the MPFC are believed to code mainly for aspects of social cogni-tion such as behavioural prediccogni-tion as a ToM precursor. While more ventral parts of the MPFC are mainly recruited by emotional aspects of confederates or interaction partners (Hynes, Baird, & Grafton, 2006; Völlm et al., 2006). The networks therefore seem to afford activity in different decision making domains, not being restricted only to social contexts. The network encodes the value of social interaction together with the value of the agent a person interacts with, constituting a rewarding experience in itself. The own behaviour with regard to the outcome for the other constitutes as much a rewarding experience as the outcome for the self and elicited neural responses to social vs non-social rewarding experiences are comparable (Ruff & Fehr, 2014). This reinforcement learning network has been shown to adapt successfully to a wide array of tasks and serves as a reliable general learning mechanism. However, this mechanism as such would perform awfully slowly in any given task and therefore specialized cortical networks for different tasks are believed to speed up the general mechanism considerably.
The functional role of the second network focusses on others’ intentions. It is compromised of the anterior cingulate cortex gyrus (ACCg), dorsal parts of the Medial Prefrontal Cortex (dMPFC) the tempoparietal junction and the superior temporal sulcus (STS) (Behrens et al., 2009). Special attention in social contexts must be payed to the dMPFC and the STS:
Some parts of dMPFC have been implicated especially in social cognition, as opposed to other types of decision making, and especially behavioural prediction as a ToM precursor. While more ventral parts of the MPFC are mainly recruited by emotional aspects of confederates or decision making partners (Hynes et al., 2006; Völlm et al., 2006) activity in the more dorsal parts reflect an aspect of human social cognition of utmost
importance: The uniquely human ability to form triadic representations of relations between another person and an object in the world. If we engage an object with someone else it is always “Him, me and the object” . We incorporate therefore our own thoughts about a state or an object alongside the thoughts, beliefs and intentions of our confederate, constituting a three part representation (Saxe, 2006). For this ability to arise, joint attention to an object is necessary, which can be triggered e.g. by gazing or calling one’s name (Kampe, Frith, & Frith, 2003). For the triadic representations to arise, an interaction with a third agent is necessary and activity is elicited through attending to similar spacial location or object of interest (Williams, Waiter, Perra, Perrett, & Whiten, 2005).
STS activity is commonly associated with the understanding of choices or actions of other agents. Those actions can be observed directly in the form of movement (Grosbras & Paus, 2005) or indirectly as the effects of a certain action are manifested (Ramnani & Miall, 2004). While the exact role of STS activity remains up to speculation it is feasible to assume that it is recruited by witnessing the choice of another agent and integrating their behaviour into our theory of their mind. Hampton, Bossaerts, and O’Doherty (2008) applied formal mathematical modelling of reinforcement learning mechanisms to a mentalizing task, demonstrating similar computational mechanisms in social as well as learning paradigms. However, the complexity of the social situation, involving direct interaction with the other agent, make the data difficult to interpret. What they found however, was that the prediction error in case of a discrepancy between expected and actual behaviour of the other could be tracked in the STS.
Another region – the Tempoparietal Junction (TPJ) - as a part of the intentional network, has been thoroughly investigated in the context of social cognition. Activity here has been shown to reflect believes and intentions of others’. The mental state of the world of another person is monitored constantly, so the activity in that region reflects the change of that state of the world for other persons, alongside with beliefs and intentions. BOLD activity in the TPJ is alleviated by story-reading about the intentions and beliefs of protagonists, whether those states may be appropriate or not (Saxe & Kanwisher,
2003). The activity however does not propagate through more descriptional approaches to the characters in the story, such as appearance, or constitution(Saxe & Powell, 2006). As well for STS and TPJ activity during ToM tasks is right lateralized.
The most successful studies in this domain focussed on particular mechanisms giving rise to social cognition. Application of model based neuroimaging methods were employed, trying to link specific aspects of social cognition to brain regions within the networks just mentioned.
1.3 The current study
The exact roles of those regions are still poorly understood and model based predictions of activity, entangling the exact computational mechanisms giving rise to a ToM are missing. This is partly due to the nature of mentalizing tasks investigated so far. Most paradigms focus on story telling or emotional inference of other agents which makes the data hard to interpret because of missing understanding of the exact computational nature of the mechanisms under consideration, partly due to a missing bridge between philosophical stances towards a ToM. As has been seen most of the prominent studies and interpretations focus rather on a Sally-Anne like paradigm affording participants to adopt a point of view of another person without decomposition of interpretable task aspects. The current work aims at bridging that gap between structural investigation of areas and networks affording social behaviour and interpreting the functions of activ-ity in these regions. Thereby the philosophical stance of a simulation theory is made possible and the dogma of earlier theory theory approaches are discarded. By shifting the focus from a conceptual idea of localized activity to a framework where investiga-tion of the role of different model parameters on activity, the decomposiinvestiga-tion of different task aspects to predict behaviour becomes possible. In comparsion to earlier studies the parameters giving rise to participants behaviour are more isolated and controllable in the current work. Furthermore the decisions depend on a discrete distribution of the amount of evidence and not merely on a probability to say yes or no in general. With the help of carefully constructed model parameters different parts of social behaviour can
be investigated in isolation. The respective change in activity depending on the change of a parameter therefore gives the possibility to investigate the role of subregions within decision networks in the human brain.
Yang and Shadlen (2007) applied such a paradigm in a non-social setting in their nowaday infamous weather prediction task study. Where monkeys made a probabilistic choice based on several symbols presented in a succession, while activity in parietal area LIP was probed. They found a representation of the actual weight with which a symbol changes the outcome in the form of a log likelihood function of the associated weight change over symbols in LIP neurons. This ’evidence accumulation framework’ led to the eventual development of the current work serving as inspiration for a quantifiable ToM task suitable for human fMRI investigation.. The current work aims to disentangle the precise roles of brain areas involved in decision making based on social and non-social information in a quantifiable manner. Using the evidence accumulation paradigms allowed us to quantify the type of information about the real state of the work (as in normal decision making) and the state of the participant (the social element). Information is formalized as red and blue dots on windows of a screen visible to the participant and a virtual actor. Each window was visible to either both players or only to the participant.Therefore the amount of red an blue dots representing either shared evidence between both players or private evidence for the participant can be quantified directly.
This manipulation allows us to differentiate the BOLD signal for representations of the state of the world of the subject as well as the representation of the confederate’s state of the world. Subjects were asked to predict the behaviour of the other based on the information available to them and received feedback on the behaviour of the other, which allowed us to see how well they can distinguish between the different types of information Subjects also had to learn the behaviour of the different confederates, since they not performed equally well on the task. This difference in confederate skill allowed us to investigate the influence of participants’ knowledge about their confederate on activity linked to evidence accumulation and prediction error. The question arising from the current paradigm is thus: How is shared information accumulated in comparison to
private information in order to form an idea about the others behaviour intentions and beliefs about their and the own state of the world?
This paradigm allows us to investigate which areas of the brain accumulate evi-dence important for making one’s own decisions, similar to that the macaques of Yang and Shadlen (2007) did, and evidence important for predicting decisions of others. We expect intraparietal areas to be important for evidence for the self, but that the TPJ will be critical for accumulating evidence relevant to others, as in ToM. However, we focus in this communication on how the evidence steers the two decisions the participant has to make: what the state of the world is and what the other thinks the state of the world is. We focused on areas on the medial frontal cortex discussed above that have been repeatedly associated with action selection, reward learning, and social decision making (Coricelli & Nagel, 2009; Rushworth, Noonan, Boorman, Walton, & Behrens, 2011). Our paradigm allows us to dissociate which regions have access to all information, which areas dissociate between privileged and non-privileged information, and how they take into account the skill of the other player. Not much is known about the exact role of dMPFC mediation forming the triadic representations, therefore the exact role of this area in the proposed network of social evidence accumulation is up to speculation. One hypothesis might be that after learning about the performance of a confederate, dMPFC activity will suppress the surprisal signal in the STS after witnessing a discrepancy between own and other’s prediction when participants know that the confederate is very bad at the task and vice versa (the signal will be alleviated in response to a discrepancy of a very clear answer when playing with a well performing avatar).
2 Methods 2.1 Participants
In total 30 participants (of whom 71.4% female, age; M = 24.76, SD = 3.65 took part in the experiment, 7 of those subjects were excluded later because they felt uncomfortable before finishing the experiment, the data showed accidental findings or there were traces of metal in the head unknown by the participant. The subjects were all right handed and
reported no history of psychological or neurological disorders and normal or corrected to normal vision. Recruitment took place through the participant pool of the Radboud University (Nijmegen, the Netherlands). Participants gave informed consent prior to the start of the experiment and could cancel or withdraw from the study voluntarily at any time without negative consequences. They received a compensation of 10 euro per hour (thus a total of 25 euro for 1 hour training and 90mins scanning) for participation. The recruitment and experimental procedures were approved by the relevant ethics authority (CMO Arnhem-Nijmegen) and in accordance with the Declaration of Helsinki.
2.2 The Task
During the whole experiment the participants watched a screen where they witnessed the ongoings from a first person point of view, facing a virtual confederate gazing at the same part of the screen where information was presented. In the centre of the screen was a black board with a dark grey square on it resembling the screen on which information was presented. In the course of a trial the colours of bits of the screen (each bit encompassing one ninth of the total screen size) changed to either white or light grey. They received the instructions that white screen resembles a window open to both players - the information there was thus visible to them as well as to their confederates - whereas the light-grey screens were omitted from the view of their confederates. Each trial consisted of 3 (12.5% of trials), 4 (75%) or 5 (12.5%) of those open screens on which clouds of blue and red dots were placed. Each screen was visible for a 1.5- 2.5s interval (normal distribution with M = 2.0s) (see Figure 1 Accumulation).
After viewing all screens participants were asked to indicate whether they had seen more red or blue dots for themselves (see Figure 1, Decision self) as well as for the confederate (see Figure 1, Decision other), by pressing a button corresponding to the position of the word red or blue on the screen. The mapping between answers and response buttons switched randomly each trial to avoid motor preparation and premature responses. After a button press was recorded the answer was framed in the respective colour.
Figure 1 . A schematic representation of the task with the 5 relevant phases,
accumulation, decision for self and other, evaluation and rating. The given times are means, with normal distributions with SD = 500ms. Only between decision self and decision other the distribution is poisson tilted from 2000ms to 6500ms with
Participants received feedback by highlighting the colour (500ms) of the actual behaviour of the other (see Figure 1, Evaluation).The behaviour of this other agent was modelled with a softmax function of evidence against probability to choose the respective answer. Since this behaviour was modelled to be imperfect the feedback not necessarily reflected the right answer, providing participants the possibility to learn about the others’ ability at the task. The inferred ability was subsequently probed every 10 trials, where participants had to judge performance on a 1 to 10 scale from “She/He is doing very bad” to “She/He is doing very good” (see Figure 1, Rating). The dot placement schedules were created randomly. In total 6 different schedules were used in order to be able to address for schedule specific effects in behaviour.
2.3 Training
Preceding the actual fMRI experiment participants completed one our of behavioural training, which was split up into 4 phases each adding an additional aspect to the task. During the first session they only saw non-occluded screens and had to indicate whether they had seen more red or more blue dots, in contrast to the later experiment they received feedback on their own choice. The second phase closely resembled the first one with the only difference being the absence of feedback, which was congruent to the later experiment. Each of those training phases consisted of 10 trials. In the third training session they were asked for to also indicate answers for another avatar, who was seen but temporarily inactive during the first two sessions. Duration of a fixation cross between questions probing for self and other decisions was normally distributed between 1.5 and 2.5s with a M = 2.0s in all training sessions. In the third phase the behaviour of the second avatar was perfect so the feedback they received on the actual behaviour of their confederate also always was the right answer from that point of view. This training phase consisted of 50 trials. The last phase had all the aspects of the true experiment in it, with two different avatars, possessing two different ability levels. In contrast to the later experiment participants were asked to rate the ability of those avatars every 5 instead of every 10 trials. The ability levels also differed from those in the final experiment to avoid
confounds through carry over learning effects from training. During the training session viewing distance was approximately 80cm on a full HD 19 inch monitor.
2.4 The Experiment
In the fMRI experiment participants were placed in a supine position (head first, lying on their back) in the scanner. There they were asked to indicate a decision concerning the total number of dots on windows of a screen between themselves and a confederate playing the task with them at the same time. In total they completed 160 trials with 4 different confederates (40 trials per confederate). Those other players possessed two different ability levels, assigned at random to avatars at different moments in the experiment. Avatar identity was randomized as well in order to avoid confounds associated with the identity and appearance of the avatars. Each trial they were asked to indicate whether they had seen more red or blue dots themselves (see 1, Decision Self) (thus referring to information shown on all screens), furthermore they had to indicate what their counterfeit should say based on the information available to them (see Figure 1, Decision other) (white screens only). To later decorrelate regressors for self and other decisions, the time between the question in the fMRI experiment followed a Poisson distribution ranging from 2 to 6.5 seconds M = 3.5s. After indicating their answer they received feedback on the actual behaviour of their confederate (see 1, Evaluation), thereby providing a possibility to learn about the ability of the other person. This ability had to be indicated on a 1-10 scale every 10 trials (see Figure 1 Rating) (4 ratings per confederate thus). During the scanning sessions participants watched a projection of the screen through a mirror placed on top of the MRI antenna. Viewing distance was adjusted for each participant individually, so that their head position in the fMRI scanner was appropriate. The scanning session per participant took approximately 90 minutes, of which one hour was the experiment and the rest preparation and a high resolution structural scan.
2.5 Data Collection
The experiment was set up using Presentation 18.3 (Neurobehavioural Systems Inc., Albany, California USA). Dot placement for the different trials, as well as
confeder-ate behaviour was constructed using custom software in Matlab 2015a (Mathworks Inc., Nattick, Massachusets, USA). The training sessions were conducted in experiment cubi-cles at the Donders Institute for Cognitive Neuroimaging (Nijmegen, The Netherlands) on a standard PC with a full HD wide screen monitor. Answers were given with the buttons “1”, “2” and “3” (for left answer, right answer and confirm rating respectively). FMRI Data was collected using a Siemens (Garching, Germany) PrismaFit 3T MRI scanner, also at the Donders Institute for Cognitive Neuroimaging. During the whole experiment Multi-Band 4 data was collected with T R = 1500ms, T E = 32 ms , voxel size = 2.0 * 2.0 * 2.0 mm. Besides that a high resolution structural scan with voxel size = 1.0 * 1.0 * 1.3 mm, T R = 2530ms and T E = 2.13ms was recorded. Participants could respond with a button box placed in their right hand, where the button mapping from left to right was in concordance with the button mapping of the behavioural training sessions.
2.6 Behavioral Data Analysis
To analyse behaviour of the participants the evidence levels (numbers of dots) where separated in different bins ranging from -35 to 35 using 5 point bins. Subsequently the average probability of choosing red over trials where the total evidence falls inside that bin was calculated. thereby we assessed participants’ tendency to choose in the face of different amounts of evidence.
We fitted different multiple linear regressions each to accuracy (with the absolutes of information), choice (the number of blue dots reflected as positive and red dots as negative numbers) and logarithm of reaction time in order to achieve a more normal dis-tribution of the otherwise poisson tilted reaction times. A first multiple linear regression consisted of regressors modelling shared information, occluded information, total number of dots, and sample length. A second multiple linear regression included separate regres-sors modelling evidence visible on each sample and regresregres-sors modelling evidence shown on the occluded samples, resulting in a total of 7 regressors. In a third multiple linear regression, regressors modelled shared information presented in the presence of a good
and a bad agent.
Each of those analysis were carried out for self and other decisions. To assess whether the influence of the regressors on the respective variable differ significantly from each other we carried out a two-way ANOVA with two factors with two levels each (Self and Other and Shared and Occluded Evidence) and mean beta weights for accuracy, choice and RT as independent variable. In order to compare mean beta weights of accu-racy choice and RT for the second model we ran a two-way ANOVA, accounting for an unbalanced design with five regressors for shared evidence and two for occluded evidence. For the last model we ran a two-way ANOVA with 3 factors with two levels each (self and other, shared and occluded evidence and good and bad Agent) comparing mean beta weights separately for accuracy choice and RT.
Agent ratings were assessed separately for agent ability and tested against each other with using a two sample t-test.
2.7 Imaging data analysis
2.7.1 Preprocessing. Preprocessing of the MRI data was carried out using tools of the fMRIB software library (www.fmrib.ox.ac.uk/fsl). For all the general Linear Model analyses described below the steps were identical. At a single subject level we carried out several steps starting with removing non brain material from the images with the Brain Extraction Tool (bet) (Smith, 2002). We spatially smoothed the data using a 5mm Gaussian Kernel at full width half maximum. Functional data was registered to the participants’ high resolution structural image using boundary based registration.
2.7.2 First-level analyses. At the single subject level a GLM with 23 de-pended or explanatory variables (EV) or regressors were entered. All EVs were normal-ized to make their range independent of their means and EVs were convolved with the hemodynamic response function. The regressors are designed to take into account the different phases of the trials. Regressors relating to the sampling of shared evidence with onset times of those samples were: Prior sum of evidence, indicating the total amount of accumulated evidence until the current sample, computed with
P rior_Sum =
current_sample−1 X
x=1
Nx (1)
where N is the amount of evidence and x the sample number. Signal update, indicated how much the total evidence changed with the current sample:
Signal_U pdate = Ncurrent_sample− Ncurrent_sample−1 (2)
with N being the total amount of evidence at that sample. The differences of Prior Sum between shared and occluded samples was calculated with
P rior_SumDif f erence = P rior_SumShared− P rior_SumOccluded (3)
. Accordingly the Signal Update Difference was calculated with
Signal_U pdateDif f erence= Signal_U pdateShared− Signal_U pdateOccluded (4)
Since embedded occluded samples will change the accumulation and update pro-cess for self and respectively other information those regressors differed after occluded samples were shown. Furthermore a binary regressor for change of the correct answer was included: Choice_Change = 1, if y−1 P x=1 Nx < 0 & y P x=1 Nx > 0. 1, if y−1 P x=1 Nx > 0 & y P x=1 Nx < 0. 0, otherwise. (5)
which was 1 if the correct answer changed as a result of the signal update presented at the current sample and 0 otherwise (before normalization). In Equation 4, y refers to the total number of samples in a trial and N to the amount of evidence, Nx is thus
the amount of evidence at sample x. The last regressor here was sample number in a trial. The same regressors were constructed for occluded evidence with the onset times of occluded samples. For the self decision phase, regressors for shared and occluded evidence were created.
Evidence = x ∗ abs(Nblue− Nred) x = 1, if answer correct. −1 otherwise
Evidence in a trial was calculated by taking the absolute of the difference between the number of red and blue dots (N ). The amount of evidence in a trial was signed according to the answer given by participants – positive for correct answers and negative for incorrect answers (x in equation 5). The same regressors were again created for other’s decision with onset times of the other relating question, evidence was signed after the same rule. For the outcome evaluation three different regressors were constructed each setting on at the beginning of the feedback participants received about the choice of the other. One regressor signalled the prediction error between the predicted answer given by the subject in the other decision phase and the actual answer given by the agent:
P rediction_Error = a − P (b|Evidence_other) a =
1, if agent answer correct. 0, otherwise.
where a refers to the answer of the agent and P is the chance to choose the correct answer (b) given the evidence. The second regressor indicated agent ability level (1= good agent, 0 = bad agent). Finally, a third regressor coded for the interaction between prediction error and agent ability level
Interaction = P rediction_Error ∗ Agent_Ability (6) We modelled the hemodynamic response function with a gamma function; this is a normalization of a gamma probability density distribution with P hase = 0, SD = 3s and M eanLag = 6s. Participant data was registered to MNI 152 2mm standart space using non-linear transformation with 10mm warp resolution (Jenkinson & Smith, 2001). Feat (Version 6.0) was used to analyze the data statistically using FILM with local autocorrelation correction (Woolrich, Ripley, Brady, & Smith, 2001).
2.7.3 Group-level analyses. Group level analysis was conducted in Feat (Version 6.0) with a FLAME 1 model using automatic outlier correction. In total 23 subjects were included in the group level analysis. All of the reported results were sig-nificant at a threshold of z = 2.3. For references the MNI 152 coordinates of the voxel inside of the significant cluster fitting the model most accurately were given after every reported result in the format: [x y z].
3 Results 3.1 Behavioural results
Behavioural data analysis revealed that participants understood the task and that they are able to behave as intended. In Figure 1 the probability for a certain choice given the amount of shared evidence can be seen. Given a high amount of evidence the chance to choose the respective answer was higher, whereas performance broke down to chance level for less certain trials (i.e. where near-equal amounts of information for both answers were presented). Besides that, occluded evidence increased the chance to answer correctly for respective self answers more than that for other answers. This was expected since occluded information should not impact the choice for the others (see Figure 2.
3.1.1 Regression on total evidence. First we ran a multiple linear regres-sion to test whether shared and occluded evidence,the number of dots, and sample length had an impact on behaviour. The first regression indicated that our behavioural ma-nipulation was successful. Shared evidence had a significantly higher impact than zero on as well self accuracy (M = 0.95, SD = 0.69), t = 6.56 , p < 0.001 and choice (M = 1.66, SD = 0.82), t = 9.66, p < 0.001). Likewise the impact of shared evi-dence on others’ accuracy (M = 1.05, SD = 0.52), t = 9.67, p < 0.001 and choice (M = 1.74, SD = 0.94), t = 8.82, p < 0.001 was significantly different from zero. Occluded information had an influence significantly higher than zero on self accuracy (M = 1.62, SD = 0.65) t = 12.05, p < 0.001 and choice (M = 2.39, SD = 1.02) t = 11.24, p < 0.001. The influence of occluded evidence on other answers’ accuracy (M = 0.27, SD = 0.47), t = 2.755, p = 0.012 and choice (M = 0.30, SD = 0.69),
Figure 2 . On the X-Axis you can see the amount of evidence with positive numbers indicating more blue than red dots on the Y-axis you can see the respective chance to choose blue according to the amount of evidence. The top row refers to shared and the bottom to occluded evidence, whil the left column shows self and the right others’ decisions.
t = 2.12, p = 0.045 was significantly higher than zero as well. The impact of occluded evidence was significantly different from zero for self accuracy (M = 1.66, SD = 0.63), p < 0.001 and choice (M = 2.48, SD = 0.95), p < 0.001. Additionally it differed sig-nificantly from zero for other’s accuracy (M = 0.20, SD = 0.35), p = 0.01 and choice (M = 0.19, SD = 0.45), p = 0.05, although far less pronounced compared to the self answers; the data showed a trend towards non significance (see Figure 3 top and middle row).
This trend indicated that participants were not able to completely isolate occluded evidence when reaching a decision for the other. The manipulation was successful, how-ever, if since the impact of occluded information is significantly higher for self compared to other decisions.
A two way ANOVA with mean beta weights of accuracy as dependent variable and evidence (shared and occluded) and answer condition (self and other) as independent 2
Figure 3 . The bar diagram shows the mean beta weights for explanatory variables (X-Axis) over all participipants on the Y-axis. The left of the panel refers to self decisions, the right to other decisions. The top row shows a regression on accuracy, the middle one on choice and the bottom row on reaction time. ∗p < 0.05, ∗ ∗ p < 0.001
level factors revealed a significant interaction effect F (1, 44) = 26.03, p < 0.001. A post hoc comparsion of means with tukey HSD corrected alpha showed the mean betas of shared evidence for self and other showed no significant difference between self and other accuracy for shared evidence (p = 0.942), in comparsion to that the difference for occluded evidence was significant p < 0.001. Thus occluded evidence influenced self accuracy significantly more than other accuracy while the influence of shared evidence was equal for both decisions.
A two way ANOVA with mean beta weights of choice as dependent variable and evidence (shared and occluded) and answer condition (self and other) as independent 2 level factors revealed a significant interaction effect F (1, 44) = 30.17, p < 0.001. A post hoc comparsion of means with Tukey HSD corrected alpha showed the mean betas of shared evidence for self and other showed no significant difference between self and other beta weights for choice for shared evidence (p = 0.09), while the difference for occluded evidence was significant p < 0.001. Occluded evidence influenced self choice
thus significantly more than others choice, while the influence of shared evidence on choice was equal in both decisions.
A two way ANOVA with mean beta weights of log(RT ) as dependent variable and evidence (shared and occluded) and answer condition (self and other) as independent 2 level factors reveal with a significant interaction effect of decision and evidence F (1, 88) = 44.22, p < 0.001. Post hoc comparisons revealed a significant difference in reaction time beta weights between shared evidence for self decisions and occluded evidence for self conditions (p < 0.001), as well as occluded evidence for other decisions (p < 0.001) (see Figure 3 bottom row).
3.1.2 Regression on single sample information. To test whether partici-pants successfully integrated all bits of information equally we ran a linear regression for the information shown on each sample and the occluded samples separately on behaviour (see 4). Again the influence of the occluded samples is significantly higher than zero for self accuracy; first occluded (M = 1.14, SD = 1.83), t = 2.99, p = 0.007 and second occluded (M = 1.62, SD = 0.82), t = 9.51, p < 0.001. Also the impact of occluded infor-mation on self choice was significantly higher than zero, with first occluded (M = 1.50, SD = 1.83), t = 3.92, p < 0.001 and second occluded (M = 2.13, SD = 0.87), t = 11.73, p < 0.001.
A significant deviation from zero is found on other accuracy for the first occluded (M = 0.83, SD = 1.38), t = 2.86, p = 0.009 but not for the second occluded sample (M = 0.23, SD = 0.62), t = 1.76, p = 0.09. The impact of occluded evidence on other choice was significantly different from zero for the first occluded sample (M = 1.20, SD = 1.71), t = 3.36, p = 0.003) and again the impact failed to reach significance for the second occluded sample (M = 0.23, SD = 0.64), t = 1.76, p = 0.09 (see Figure ??, top and middle row).
Figure 4 . The bar diagram shows the mean beta weights over all participants (Y-Axis) for explanatory variables (Evidence on each sample) on the X-Axis. The left of the panel refers to self decisions, the right to other decisions. The top row shows a
regression on accuracy, the middle one on choice and the bottom row on reaction time. ∗p < 0.05, ∗ ∗ p < 0.001
A two way ANOVA for unbalanced designs revealed a significant interaction effect between shared and occluded evidence and self and other decisions for the mean beta values of accuracy, F (1, 318) = 27.17, p < 0.001. Post hoc Tukey-HSD corrected compar-isons showed no significant difference in the mean beta weights for shared evidence on accuracy for self or other decisions (p = 0.15) while the mean beta weights of occluded evidence were significantly higher for self than for other decisions (p < 0.001).
A two-way ANOVA for unbalanced designs revealed a significant interaction effect between occluded and shared evidence and self and other decisions for the mean beta val-ues of choice, F (1, 318) = 52.55, p < 0.001. Post hoc Tukey-HSD corrected comparisons showed no significant difference in the mean beta weights for shared evidence on choice for self or other decisions (p = 0.97) while the mean beta weights of occluded evidence were significantly higher for self than for other decisions (p < 0.001).
between occluded and shared evidence and self and other decisions for the mean beta values of log(RT) F (1, 318) = 4.8, p < 0.001 (see Figure 4, bottom row).
3.1.3 Regression on evidence split by agent ability. Lastly we ran a multiple regression with different regressors for evidence shown at each agent ability level in order to see whether participants weighted information differently depending on the the ability of the agent. Influence of occluded evidence on self accuracy for the good (M = 1.61 SD = 0.77), t = 10.03, p < 0.001 as well as for the bad agent (M = 1.57, SD = 0.58), t = 13.01, p < 0.001 was siginifcantly higher than zero. The influence of occluded evidence of the good (M = 2.29, SD = 1.02), t = 10.77, p < 0.001 and bad agent (M = 2.51, SD = 1.12), t = 10.13, p < 0.001 was significantly higher than zero for self choice. Other accuracy however was not influenced significantly by occluded evidence of the good agent (M = 0.18, SD = 0.70), t = 1.27, p = 0.22 in contrast to occluded evidence of the bad agent (M = 0.28, SD = 0.44), t = 3.13, p = 0.005. The influence of occluded evidence on other choice was significantly higher than zero for the good agent (M = 0.30, SD = 0.62), t = 2.29, p = 0.03, while the difference was not significant for the bad agent (M = 0.32, SD = 0.79), t = 1.92, p = 0.07. This is a strange picture, what is truly interesting here is whether the influences of occluded evidence for the good and bad agent differ significantly, since this would tell something about the actual importance people assign to the evidence, granting the good agent more knowledge than possible. We further investigated this finding by means of a three-way ANOVA in the following paragraph.
Figure 5 . The bar diagram shows the mean beta weights over all participants (Y-Axis) for explanatory variables (Evidence split by agent ability) on the X-Axis. The left of the panel refers to self decisions, the right to other decisions. The top row shows a regression on accuracy, the middle one on choice and the bottom row on reaction time. ∗p < 0.05, ∗ ∗ p < 0.001
A three-way ANOVA revelead a significant interaction effect between self and other and shared or occluded evidence F (1, 177) = 55.11, p < 0.001 on accuracy. Post-hoc comparisons with Tukey-HSD revealed a significant difference in beta weights for occluded information in self compared to other decisions (p < 0.001). All other interactions were not significant, also a main effect for agent ability failed to reach significance, indicating that participants indeed treated agents as equal with respect to the knowledge they imposed upon them.
A 3-way ANOVA revelead only a significant interaction effect between self and other and shared or occluded evidence F (1, 177) = 61.05, p < 0.001 on choice. Post hoc comparsions with Tukey HSD revealed a significant difference in beta weights for occluded information in self compared to other decisions (p < 0.001). No other interactions were not significant, also a main effect for agent ability failed to reach significance, indicating that participants indeed treated agents as equal with respect to the knowledge they
imposed upon them.
A three-way ANOVA with logRT as dependend variable and agent ability (good; bad), decision (self; other) and evidence (shared; occluded) as three independent two level factors demonstrated a main effect of decision F = (1, 177) = 67.42, p < 0.001 as well as a significant main effect for evidence F (1, 177) = 5.43, p = 0.02.
The fact that information shown on sample 5 has such a big impact is probably due to the small amount of 5 sample trials (12.5%) and therefore in the case of presence the impact is comparably high. Another feasible explanation could also be that the longer trials are just considerably more complicated.
3.1.4 Agent rating. A two sample t-test of the agent rating scores revealed accurate separation between the two agent ability levels as shown by a significant differ-ence between the good and the bad agent, t = −10.58, p < 0.001 (see Figure ??).
Figure 6 . The top graphic shows the mean rating over all trials (Y-Axis) per participant (X-Axis). The bottom left shows the rating (Y-axis)over single trails (X-axis) averaged over participants for the two ability levels, while the bottom right shows the average rating (Y-axis) over single rating trials (X-axis) for all avatars individually.
In summary, we conclude that for all regressions, stronger evidence led to higher accuracy and more evidence towards a certain decision resulted in choosing the respective colour. Furthermore, shared evidence had an equal impact for self and other answers in all three models while the impact of occluded evidence was significantly higher for self than for other decisions. Reaction times showed that more evidence makes responses faster, but only for the self decisions. This connection may have been mediated by certainty. The more certain someone was of her answer the faster she will give it. The absence of an effect of evidence on reaction time for other decisions was probably due to the long lag between decisions; by the time participants indicate their decision it is already made so behaviourally no significant effects could be witnessed. Finally, participants were able to distinguish the agent ability levels and learned to judge an agent quickly and reliably.
3.2 fMRI Results
We concentrate in this report on the decision, evaluation and rating phases of the trials and leave an analysis of the accumulation phases for later reports (see Figure 1). Also, we concentrate here on activations in the frontal lobe, leaving a more detailed analyses of the involvement of temporoparietal cortex for future reports.
We found that different parts of the medial frontal cortex respond differently during the decision phases. Shared evidence during self decisions showed activation in the ventromedial prefrontal cortex (vmPFC) [-4 48 -4] and dorsal ACC (dACC) [48 66 68]. In contrast, occluded evidence during self decisions activated the paracingulate cortex [8 20 38]. This indicates that these regions, although all playing a role in making the self decisions, have access to different types of information, with only paracingulate distinguishing between information that is priviledged to the subject.
Shared evidence during other decisions activated the vmPFC [6 34 -14], overlap-ping ventrally with the activation found for shared evidence in self decisions as well as the dACC [0 -2 38]. Activation for occluded evidence in other decisions in contrast activated the pre-supplementray motor area (pre-SMA) [4 16 58].
dACC also tracked shared evidence for both decisions but only occluded evidence for self decisions and the pre-SMA tracked occluded evidence only for other decisions.
Figure 7 . Saggital and Coronal view of the brain with the mean results over
participants for the different types of Evidence, Analysis in the paper is restricted to circled areas. All results are significant at a z = 2.3 threshold
We manipulated the skill level of the agents; two agents performed well, two others badly. Since we expected participants to learn about the ability of those agents, it is reasonable to expect differences in activation due to agent ability. After the decision phase, the actual choice made by the agent was communicated to the participant, allowing evaluation of their behaviour.The ability of an agent gave a signal in the frontal pole [38 58 2], indicating monitoring of the current agent ability. A prediction error indicating the difference between the subject’s response during the other decision and the actual behaviour of the agent, was present in the paracingulate gyrus [8 14 52]. This signal can be used to track the ability of the agent, which itself was represented in the lateral part of the frontal pole. The ability of the agent can in turn be taken into account when evaluating their behaviour. Therefore, we searched for areas showing an interaction between the prediction error and the agent ability level. We indeed found such an interaction in the same paracingulate region [-2 12 52], showing activation tracking the prediction error.
Figure 8 . Coronal and axial view of the brain with the mean results over participants for the evaluation phase, Analysis in the paper is restricted to circled areas. All results are significant at a z = 2.3 threshold
4 Discussion
We set out to understand the neural basis of decision making in a social and non-social context by using a well established evidence accumulation paradigm that allowed us to elucidate the type of information represented by brain regions involved in solving the task. Behavioural evidence suggests that participants solved the task as intended. When making decision for themselves about the state of the world, participants took into account all presented information, while when making decisions for others, they discounted evidence that was not seen by the other agent. We tested social behaviour in a very artificial lab environment, barely resembling natural social interaction because of the necessity to carefully control stimuli. This necessity poses a considerable conceptual threat to the validity of the experiments, since the imposed environment conflicts with the complexity of natural social interactions. Unfortunately this is a pitfall most social neuroscience experiments succumb to. Therefore, research in the field of social cognitive neuroscience often faces a supposed threat to validity. The question arising then is: Is a
social experiment social after all?
This threat to validity could be counter argued by what the philosopher Daniel ? calls the intentional stance: Following this line of thought the best we can do is to project our inferences onto the actual intentions and beliefs of other agents. Hereby we rationalize our thoughts and attribute an intention to that other agent, making it possible to infer certain inner states to oneself, other individuals, animals, non animate objects or even phenomena (?). Considering this wide array of things and beings humans tend to anthropomorphize, it is reasonable to assume that the same processes recruited by social interactions are also active while participating in an experiment that tries to imitate a certain aspect of a real social interaction. Or in short: “If we infer intentions and inner states to things and people alike, does it really matter whether an experiment is social after all?“. Adopting this philosophical point of view makes a strong claim for the validity of social neuroscience and therefore lay the foundation for the validity of findings stemming from computational neuroscience focussing on social situations. Accepting this philosophical stance secures the validity of the findings discussed in the following sections.
4.1 Imaging results vmPFC and ACC
In what follows the differences in activity between self and other decisions for shared and occluded evidence are laid out. The vmPFC, as well as the cingulate and paracingulate gyrus were active in overlapping parts for self and other shared evidence.
We found vmPFC activity for shared information in both decisions. Currently, the views revolving around the vmPFC’s role in decision making as triadic representations lack any kind of quantitative interpretation but support the idea that the relationship between the self, others and an object in the world is represented through activity in that region (Baron-Cohen, 2005). The current study gathered additional evidence in favour of this view because shared evidence in a trial can be regarded as the object representation shared amongst both players. Furthermore studies investigating value based decision making plead for vmPFC activation scaled by choice value (Rushworth et al., 2011). While the current paradigm lacks any value of a decision - because the answers
are not rewarding or punishing per se - it can be argued that the amount of evidence translates to the value of a certain choice. Another interpretation of choice value called “other regarding preferences” hypothesizes that the vmPFC is not at all concerned about any kind of social stimuli but codes choice value according to the altruistic tendency of participants. From this point of view a correct decision for another agent and the subsequent guiding and help for that agent is a rewarding experience per se. Therefore this choice value is also coded in the vmPFC,rendering activity in that area as purely value-based and non-social (Cooper, 2010). However, since the prediction of the other agent’s behaviour did not influence the behaviour of the agent in subsequent trials, nor did it help him in any way she could benefit from a correct prediction, this interpretation of vmPFC signal can be ruled out with the current paradigm for social decision making, regardless of individual differences in altruistic tendencies.
We found activation in the dACC scaled by shared evidence for the self and for the other decisions, as well as for occluded evidence for the self decisions. The fact that dACC also codes occluded evidence distinguished it from the vmPFC that only codes shared evidence. The dACC therefore only coded evidence that is important for a decision. The exact role of dACC in decision making is a matter of active debate. The dACC’s role is thus attending to and monitoring of task relevant information in non-social paradigms Rushworth et al. (2011). This interpretation is consistent with the current findings. The separation of activity for self and other decisions according to occlusion support this hypothesis: The signal in the dACC was only scaled by information directly relevant to the decision at hand.
Occluded evidence for the self distinguishes itself from other kinds of evidence by an activation of the paracingulate gyrus, while occluded evidence for the other only afforded activity in the preSMA (see next paragraph). Paracingulate thus seems to have a role in coding privileged information used in the current decision. In contrast, activity scaled by priviledged information not used for the decision at hand is found in the more dorsal pre supplementary motor area (pre-SMA) as is explained in the next paragraph.
pre-SMA. Nachev, Kennard, and Husain (2008) showed that this region is involved in the inhibition of decisions and actions, but it has not been shown in any social context so far. Interestingly, the negative activation of pre-SMA with occluded evidence suggests that it has a role on preventing the actor to act on the occluded information, which is exactly what is needed to solve the other decision. This result further challenges the idea that parts of the frontal cortex are uniquely involved in social cognition. It rather supports the idea of a domain-general mechanism solving problems using similar computations independent of the information’s relation to the self or others. Interestingly, while pre-SMA’s correlation with occluded evidence was negative, a positive correlation was found in the lateral temporoparietal junction area (TPJ) result not reported -an area often associated with uniquely social information processing (Saxe, 2006). One possible interpretation of this combination of effects is that the evidence accumulation for occluded evidence-i.e., evidence that dissociates between the subject’s knowledge of the world and the other’s model of the world- is coded in the TPJ, consistent with its proposed role in ToM, and that this information is then used by pre-SMA to reach the correct decision. This view also converges with pre-SMA’s proposed role in adjustment of ongoing behavioural programs based on relevant environmental information (Mars et al., 2009). The TPJ-coded net evidence amount would here serve as an indication for the amount of adjustment needed to ultimately infer the decision for the other. This possibility will be explored in future additional analyses of the data.
Apart from requiring the subject to represent the other’s model of the world, the experimental paradigm also manipulated the ability level of the other agents. Knowledge about the abilty of others to correctly integrate information is crucial to successful in-teractions. Participants received feedback on the choice of the other on each trial. We found this information affected activity in a set of prefrontal regions (see Figure 8).
A participant could learn about the other’s ability by comparing their estimate of what they should do (’other’ decision) with the actual choice made. The difference between these two signals is a prediction error that we found in the paracingulate cortex in a region partly overlapping with the dACC - hotpot for occluded information. Over
time, these prediction errors allow one to build a representation of the ability of the other. We found that this signal was represented in the lateral part of the frontal pole (FPl). Interestingly, this region is often reported in decision making tasks as coding for information that is important overall, but not necessarily relevant to the trial or decision at hand. For instance, in a decision making task featuring three options where you are forced to choose between two of them, the third irrelevant option is represented in FPl (Boorman, Behrens, Woolrich, & Rushworth, 2009). Consistent with our observation that FPl codes the long-term ability of the other’s ability, Summerfield and Koechlin (2009) emphasized FPl’s role in integrating information over long time scales. Here, we show that this role extends to social information (compare (Hartogsveld et al., 2017)). Interestingly, dACC itself showed not only the prediction error, but also an interaction between the prediction error and the ability level of the agent. This might indicate that these two regions interact in determining the ability of the agent, but a full investigation of their interaction is left to the future. Alternatively the interaction could be interpreted as an adjustment of the magnitude of the frontal pole signal according to the prediction error, affording less surprisal on the side of the participants when witnessing a correct answer in less clear trials.
Participants’ behaviour was modulated in an interesting way by the agent abil-ity (see Figure 5 top right). The final regression with which the behavioural data was analysed shows the difference in impact of evidence in the presence of a good or bad avatar. The most striking finding here demonstrates that participants did not integrate the occluded evidence to arrive at a decision in the presence of a good avatar, while they did so in the presence of a bad avatar. This demonstrates a better and more stringent isolation of occluded information n the presence of a good compared to a bad avatar. The exact mechanisms of brain activity leading to that interesting result lies in the realm of later connectivity analysis to see what kind of communication pattern leads to such an unexpected outcome. It is feasible to assume, that the presence of a better avatar gives some stronger signal which in turn inhibits evidence accumulation for occluded evidence more strongly as compared to the worse avatar.
In conclusion, the data reported here give a first overview over a mechanistic understanding of the brain in social situations. We showed that in order to arrive at a decision in a social context, not only commonly termed ’social’ brain areas, such as TPJ or STS are recruited to solve a task, but areas commonly seen in single-person decision making tasks draw on social information in order to compute a decision. The idea of a thoroughly social brain is thus further challenged, especially in the case of the frontal cortex. Our results are thus in line with theories arguing for more general decision making mechanisms that use specific computational strategies as input (Vassena, Holroyd, & Alexander, 2017); in the current work this information happens to be of social nature. This approach opens the door to a much more computationally grounded understanding of the neural basis of social decision making.
5 References
Baron-Cohen, S. (2005). The empathizing system. Origins of the social mind: Evolu-tionary psychology and child development, 468–492.
Behrens, T. E., Hunt, L. T., & Rushworth, M. F. (2009). The computation of social behavior. science, 324 (5931), 1160–1164.
Boorman, E. D., Behrens, T. E., Woolrich, M. W., & Rushworth, M. F. (2009). How green is the grass on the other side? frontopolar cortex and the evidence in favor of alternative courses of action. Neuron, 62 (5), 733–743.
Cooper, R. P. (2010, oct). Cognitive Control: Componential or Emergent? Topics in Cognitive Science, 2 (4), 598–613. Retrieved from http://doi.wiley.com/10 .1111/j.1756-8765.2010.01110.x doi: 10.1111/j.1756-8765.2010.01110.x
Coricelli, G., & Nagel, R. (2009). Neural correlates of depth of strategic reasoning in medial prefrontal cortex. Proceedings of the National Academy of Sciences, 106 (23), 9163–9168.
Goldman, A. I. (1992). In defense of the simulation theory. Mind & Language, 7 (1-2), 104–119.
Goldman, A. I., & Others. (2012). Theory of mind. The Oxford handbook of philosophy of cognitive science, 402–424.
Gordon, R. M. (1986). Folk psychology as simulation. Mind & Language, 1 (2), 158–171. Grosbras, M.-H., & Paus, T. (2005). Brain networks involved in viewing angry hands or
faces. Cerebral Cortex, 16 (8), 1087–1096.
Hampton, A. N., Bossaerts, P., & O’Doherty, J. P. (2008). Neural correlates of mentalizing-related computations during strategic interactions in humans. Pro-ceedings of the National Academy of Sciences, 105 (18), 6741–6746.
Hartogsveld, B., Bramson, B., Vijayakumar, S., van Campen, A. D., Marques, J. P., Roelofs, K., . . . Mars, R. B. (2017). Lateral frontal pole and relational processing: Activation patterns and connectivity profile. Behavioural Brain Research.
Hill, R. A., & Dunbar, R. I. M. (2003). Social network size in humans. Human nature, 14 (1), 53–72.
Hynes, C. A., Baird, A. A., & Grafton, S. T. (2006). Differential role of the orbital frontal lobe in emotional versus cognitive perspective-taking. Neuropsychologia, 44 (3), 374–383.
Iacoboni, M., Molnar-Szakacs, I., Gallese, V., Buccino, G., Mazziotta, J. C., & Rizzolatti, G. (2005). Grasping the intentions of others with one’s own mirror neuron system. PLoS biology, 3 (3), e79.
Jenkinson, M., & Smith, S. (2001). A global optimisation method for robust affine registration of brain images. Medical image analysis, 5 (2), 143–156.
Kampe, K. K., Frith, C. D., & Frith, U. (2003). “hey john”: signals conveying communica-tive intention toward the self activate brain regions associated with “mentalizing,” regardless of modality. Journal of Neuroscience, 23 (12), 5258–5263.
Lütcke, H., & Frahm, J. (2007). Lateralized anterior cingulate function during error processing and conflict monitoring as revealed by high-resolution fmri. Cerebral Cortex, 18 (3), 508–515.
Mars, R. B., Klein, M. C., Neubert, F.-X., Olivier, E., Buch, E. R., Boorman, E. D., & Rushworth, M. F. S. (2009). Short-Latency Influence of Medial Frontal Cortex on Primary Motor Cortex during Action Selection under Conflict. Journal of Neu-roscience, 29 (21). Retrieved from http://www.jneurosci.org/content/29/21/ 6926.short
Nachev, P., Kennard, C., & Husain, M. (2008). Functional role of the supplementary and pre-supplementary motor areas. Nature Reviews Neuroscience. doi: 10.1038/ nrn2478
Premack, D., & Woodruff, G. (1978). Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1 (4), 515–526.
Ramnani, N., & Miall, R. C. (2004). A system in the human brain for predicting the actions of others. Nature neuroscience, 7 (1), 85–90.
Rizzolatti, G., & Craighero, L. (2004). The mirror-neuron system. Annu. Rev. Neurosci., 27 , 169–192.
