Dorsal posterior cingulate cortex encodes the informational value of feedback in human–computer interaction

In communication between humans as well as in human–computer interaction, feedback is ubiquitous. It is essential for keeping up the dialogue between interaction partners, evaluating the adequacy of an action, or improving task performance. While the neuroscientific view on feedback has largely focused on its function as reward, more general definitions also emphasise its function as information about aspects of one’s task performance. Using fMRI in a computer-controlled auditory categorisation task, we studied the neural correlates of the informational value of computer-given feedback independent of reward. Feedback about the correctness of a decision, compared with feedback only indicating the registration of a decision, increases activation of the dorsal posterior cingulate cortex, supporting this region’s role in adapting to behaviourally relevant information. Both conditions elicit equally strong activation of the dorsal striatum which does not support an interpretation of feedback information as a type of reward. Instead, we suggest that it reflects a more fundamental aspect of human interaction behaviour, namely the establishment of a state that enables us to continue with the next step of the interaction.


In communication between humans as well as in human-computer interaction, feedback is ubiquitous.
It is essential for keeping up the dialogue between interaction partners, evaluating the adequacy of an action, or improving task performance. While the neuroscientific view on feedback has largely focused on its function as reward, more general definitions also emphasise its function as information about aspects of one's task performance. Using fMRI in a computer-controlled auditory categorisation task, we studied the neural correlates of the informational value of computer-given feedback independent of reward. Feedback about the correctness of a decision, compared with feedback only indicating the registration of a decision, increases activation of the dorsal posterior cingulate cortex, supporting this region's role in adapting to behaviourally relevant information. Both conditions elicit equally strong activation of the dorsal striatum which does not support an interpretation of feedback information as a type of reward. Instead, we suggest that it reflects a more fundamental aspect of human interaction behaviour, namely the establishment of a state that enables us to continue with the next step of the interaction.
Feedback is essential in human dialogues as well as for a successful communication in human-computer interactions (HCI) 1 . It is of particular importance in situations where humans are trying to improve their performance in a given task, which is the case in typical school scenarios as well as when someone is trying to excel in a computer game or interacting with an intelligent tutoring system. Even though a lot is known about the effects of feedback in such contexts from decades of behavioural research [2][3][4][5][6] , the underlying mechanisms of successful feedback application are vastly unexplained 5 .
More recently, neuroimaging studies have contributed to the understanding of these mechanisms by investigating how different functions of computer feedback are represented in the brain. This includes addressing the neuronal effects of feedback timing and reliability. While delayed and omitted computer feedback elicit behavioural and psychophysiological responses that reflect irritation and frustration of the user 7,8 , fMRI studies have shown that this is accompanied by a strong activation of a brain network for attention and action control 9,10 . Despite the fact that the feedback employed in these studies only registered the participants' button press, differential activation for immediate, delayed, and omitted feedback was also found in brain areas typically associated with reward processing, e.g. the dorsal striatum 9 (see also 11 for a comparable study that used no feedback and non-contingent feedback as control conditions). Moreover, the activation differences resemble findings on temporal prediction error coding in monkeys 12 and humans 13 that, however, explicitly used rewarding feedback stimuli.
At first glance, these results may therefore appear to be in line with traditional reinforcement-based theories 14,15 that inextricably linked feedback with reward (see 16 , for a historical perspective). Following this line of thinking, the vast majority of studies on the neural correlates of feedback processing have been designed as reinforcement experiments using reward and punishment, and as a consequence, any activation in dopaminergic brain regions has almost exclusively been attributed to mechanisms of reward processing 17,18 . In this context, the dorsal striatum has specifically been associated with the anticipation of rewarding feedback 19,20 and the establishment of associations or contingencies between stimulus, response and reward [21][22][23] .

Results
In the current experiment, participants were asked to categorise frequency-modulated tones according to their modulation direction by pressing one of two buttons for rising versus falling tones. In different blocks of the experiment, each participant received feedback that was (a) registering or evaluative, and (b) visual, auditory, or verbal. In the following, behavioural and neuroimaging results are summarised.  Fig. 1). The evaluative feedback in all three modalities elicited a substantially higher BOLD response than the registering feedback, especially in the later phase of the blocks, where the BOLD response decreased almost back to baseline in the registering feedback conditions while it remained at a considerably higher plateau in the evaluative feedback conditions. No other brain regions showed significant differences in neural activity in response to evaluative feedback compared to registering feedback. Specifically, no effects of informational content were apparent in the dorsal striatum or other areas associated with the human reward system. To substantiate this, we calculated a more sensitive general linear model on a region-of-interest level for all voxels covering the dorsal striatum area previously reported to be differentially activated in the same task by immediate, delayed and omitted registering feedback 9 (Talairach coordinates: x = − 24, y = 9, z = 3; volume: 20 mm 3 ). This analysis confirmed the lack of differential feedback effects in this area (mean t = 0.94, mean p = 0.372), as did a corresponding analysis of the dorsal striatum area as identified by Behne et al. 11 (Talairach coordinates: x = − 20, y = 6, z = − 2; volume: 760 mm 3 ; mean t = 0.78, mean p = 0.498). Illustrating this lack of differences, Fig. 2 shows the striatal activation in the current study for all feedback conditions versus rest/fixation (Talairach coordinates: x = − 22, y = 1, z = 9) and the equal BOLD signal increase that is elicited in this area throughout the entire task block by both evaluative and registering feedback.

Discussion
The experiment described here aimed at distinguishing different types of feedback information in human computer interaction and their neural correlates. A particular focus was set on the role of dopamine-related brain areas, specifically the dorsal striatum, which in previous studies 9, 11 has demonstrated an increase in activation following correlated registering feedback that did not include any reward in the classical sense. To this avail, the neural responses to feedback with mere registering information were compared to the responses to feedback with additional evaluative information.
From a reward-centric perspective on feedback information in human-computer dialogues, i.e. building on the assumption that information itself has some reinforcing/rewarding properties and that more information would equal a higher reward, a stronger increase in activation in the dorsal striatum would have been expected in blocks with evaluative feedback compared to blocks with merely registering feedback. This, however, was not reflected in our results. Instead, we observed differential activation for evaluative versus registering feedback only in the posterior cingulate cortex (PCC), an area not typically referred to in studies on reward-related processing. In the dorsal striatum, on the other hand, we observed an equal increase in activation for both registering and evaluative feedback. This supports an understanding of the processing of feedback information as essentially independent from reward processing mechanisms.
In accordance with the interpretation given by Behne et al. 11 , the increase in activation that we observed in the dorsal striatum may well reflect a more basic process like the confirmation of a successful continuation of an ongoing interaction, something that is of fundamental importance especially in dialogues between humans and computers, where interruptions of communication commonly occur due to technical failure. Crucially, in the present experiment, both registering feedback and evaluative feedback were administered in a timely, correlated fashion, so that both types of feedback equally provided this essential information and, as a consequence, activated the dorsal striatum to an equal extent. www.nature.com/scientificreports/ It is important to note that we do not mean to reinterpret previous findings of striatal activation in experimental situations where feedback in fact primarily fulfilled its function as a motivator/reinforcer 17,[19][20][21]35 . Indeed, two previous studies showed that the addition of monetary rewards increased striatal activation as compared with purely evaluative feedback 36,37 . The vast majority of studies that reported striatal activation in response to feedback, however, inseparably coupled positive evaluative feedback with the prospect of monetary gain or other secondary reinforcers (e.g. 22,[38][39][40][41][42][43]. When such reinforcement is not implied and feedback is stripped down to its informational content, however, we suggest that the remaining striatal activation does not reflect any rewardrelated processes but simply indicates the identification of a functioning dialogue, as was the case with both types of informative feedback in the current study. An alternative interpretation of the equal striatal activation observed for all feedback conditions cannot be ruled out, since the accuracy in the employed categorisation task was very high: If information about successful performance (i.e., positive evaluative feedback) is perceived as rewarding/motivating to attain a performance goal 44 , one could argue that we observed striatal activation (a) in blocks with evaluative feedback since this consisted predominantly of positive feedback, and (b) also in blocks with registering feedback due to the high probability of the registered response being correct. While it is unclear to what degree participants were aware of this high probability of success, future experiments should manipulate task difficulty to address this possibility.
Beyond the discussion about the involvement of the reward system, we found that information about the adequacy of the user's task performance, which was provided by evaluative but not by registering feedback, elicited an increase in activation in a region that has previously not been in the centre of attention in the area of feedback research. Based on its Talairach coordinates, this activation conforms to BA 23 and can therefore be considered as part of the posterior cingulate cortex (PCC). As of today, the functional role of the PCC is not yet clear [45][46][47] . First and foremost, the PCC is known as one of the most metabolically active brain regions at rest and has been set at the centre of the default mode network (DMN), which shows conjoint deactivation during the engagement in a wide range of attention-demanding, externally-oriented, goal-directed cognitive tasks [48][49][50][51][52][53] .
In contrast to the task-induced deactivation typically associated with the PCC as part of the DMN, however, the cluster of PCC activation in the current study showed a pronounced positive deflection of the hemodynamic response function during the auditory categorisation task (see Fig. 1). To understand this ostensible contradiction, one needs to take into account that the DMN was originally defined based on low resolution PET data and early fMRI data using large smoothing kernels, which led to a treatment of the entire PCC as a uniform component of the DMN. In recent years, however, it has become clear that a more detailed parcellation of the PCC is required anatomically 54,55 as well as functionally 46,47,56,57 . Building on Brodmann's 58 identification of cytoarchitecturally defined areas, Vogt and colleagues 54,55 provided a novel anatomical parcellation scheme of the posterior cingulate cortex, dividing this brain area into a dorsal portion (dPCC, including the dorsal parts of BA 23 and BA 31) and a ventral portion (vPCC, including the ventral parts of BA 23 and BA 31). The resulting PCC sub-components (d23, d31, v23, v31) are illustrated in Fig. 3A.
Following Vogt and colleagues' parcellation scheme 54,55 , the activation observed in the current experiment can be attributed to the dorsal part of BA 23 corresponding to PCC area d23 in Fig. 3A. To verify that the observation of a positive BOLD response of this area in blocks with evaluative feedback does not automatically imply a conjoint task-positive activation of the rest of the PCC/DMN, we performed an additional analysis contrasting the activation in blocks of task performance (with evaluative or registering feedback) with the activation in resting blocks. This did indeed reveal the DMN-typical deactivation during task performance (Fig. 3B). The affected PCC areas, however, were either located more superior or more posterior than the observed cluster of positive activation for evaluative versus registering feedback (cf. Fig. 1), thus conforming to areas d31 and v31/v23 of the PCC.
From this pattern of activation and deactivation in different sub-regions of the PCC it becomes clear that the cognitive processes underlying the observed activation in area d23 cannot be understood by reflecting on PCC function as an undifferentiated whole. On the basis of the anatomical organisation put forward by Vogt and www.nature.com/scientificreports/ colleagues 54,55 , Leech and Sharp 46 reappraised the functional organisation of the PCC, specifically differentiating between the functions of the ventral and the dorsal portion of the PCC. In this framework, the so-called ABBA (Arousal, Balance and Breadth of Attention) model, activity of the ventral portion of the PCC is primarily associated with internally directed attention (see also 47,54,59 ), which has also been advocated as one of the processes underlying DMN activation in general 52,60-63 . Activation of the dorsal portion of the PCC, on the other hand, is primarily associated with a broad external focus of attention, leading the authors to suggest an involvement in "detecting and responding to environmental events that may require a change in behavior" 46 (p. 24). Supporting evidence for the assumption that PCC sub-regions are involved in distinct cognitive processes and do not simply function as a unitary, undivided module of the DMN stems from functional connectivity analyses of fMRI data. For example, Leech et al. 56 demonstrated that a strong functional connectivity of the PCC to the rest of the DMN is mainly exhibited by ventral portions of the PCC. The dorsal PCC, on the other hand, is functionally highly connected to frontoparietal networks of attention and executive control, suggesting a more prominent role of dPCC in the control of cognition 56 . Similarly, Fan et al. 57 reported a higher connectivity of dPCC to the central-executive network (CEN) than to the DMN (as opposed to vPCC's stronger connection to the DMN). Taken together, these observations strongly emphasise the need to consider more fine-grained specialisations of PCC subareas when discussing PCC function.
The present finding of an increased activity of the dorsal PCC (specifically area d23) during blocks with evaluative feedback is in accordance with Leech and Sharp's 46 conception of the dorsal PCC as involved in the behavioural adaptation to information from the environment. Based on its higher informational content, evaluative feedback conceivably possesses a higher relevance for behavioural adaptation than registering feedback. As a consequence, the observed increase of activity in area d23 may reflect the organism's adjustment to the available information about the adequacy of the individual's behaviour.
In this regard, it needs to be acknowledged that the current findings do not entirely support a functional dichotomisation between dPCC and vPCC as put forward by Leech and Sharp 46 , Bzdok et al. 47 , or Fan et al. 57 . Based on the differential activation and deactivation pattern in area d23 versus areas v23, d31, and v31 (see Fig. 3B), the current findings suggest an even closer correlation between the functional specialisation of the PCC and its anatomical parcellation as put forward by Vogt 55 , and thus a separate consideration of all four subregions of the PCC (cf. Fig. 3A). More precisely, such a more intricate differentiation could address the following mismatch between Leech and Sharp's 46 conceptualisation of the ABBA model and our data: While Leech and Sharp 46 indeed predict an increase in neural activity in the dPCC for a broad external focus of attention, they associate a narrow external focus with deactivation of both vPCC and dPCC. The task employed in the current study, however, must primarily be considered as requiring such a narrow external focus of attention (attending task-related feedback from the immediate environment, i.e. the technical system). Yet we demonstrated deactivation for all sub-regions except area d23, which showed the described positive deflection in the BOLD response. This result suggests to incorporate a further distinction between areas d23 and d31 within the dorsal PCC in functional conceptualisations of the PCC such as the ABBA model.
With the conceived role of area d23 in behavioural adaptation in mind, the question arises why PCC effects such as the ones observed here have not been reported earlier in the feedback processing literature. Depending on the particular study at hand, this gap may be attributed to one or more of several reasons. For example, many reward-related studies did not include any neutral or registering feedback in their experimental setup 24,26,37,39,41,64,65 or neglected contrasting it to evaluative feedback directly in favour of other comparisons more imperative to the respective publication 23,66 . Others did investigate the neural correlates of neutral versus evaluative feedback but selectively focused on specific volumes of interest like striatal regions, excluding the PCC from analysis 40 . One previous study 44 reporting whole-brain effects besides the targeted differences in the striatum did in fact list an increased activation of PCC region BA23 in a comparison of evaluative feedback with a 'no-feedback' condition (where a pound sign was shown after both correct and incorrect responses). However, since the study's focus was on the striatal activation differences, the observed involvement of the PCC was not discussed any further. Finally, many previous studies on feedback processing, even if they did implement neutral feedback in a direct comparison with positive and negative feedback (associated with monetary gain and loss), set up experimental contexts where the evaluative feedback was not contingent upon the participants' actual responses (e.g. in gambling-like tasks like the card-guessing paradigm) 22,38,39,41 . Therefore, this type of feedback did not contain information that could be utilised for monitoring task performance, which may be the reason why these studies did not observe dorsal PCC activation.
In summary, our results do not support the hypothesis that the informational value of (computer) feedback merely constitutes some kind of reward, because then more informative feedback should have led to increased activation of dopaminergic brain regions such as the dorsal striatum. Instead, the current results show that, independent of their informational content, the two types of feedback activate the dorsal striatum to an equal extent which might reflect a more basic process indicating the successful continuation of an interaction. This does not rule out the possibility of additional modulations of striatal activation when feedback is combined with primary or secondary reinforcers or when it signals a step towards attaining a rewarding outcome. The informational value of feedback, on the other hand, appears to be independently encoded in the dorsal portion of the posterior cingulate cortex (area d23), which responds more strongly to the additional information contained in evaluative feedback as compared to registering feedback. This supports the role of this brain region in the processing of information that is potentially relevant for behavioural adjustments.
Considering the importance of the informational value of feedback suggested by the present results, a more comprehensive examination of informative feedback would be conducive to attain a thorough understanding of the complex neuronal differentiation between feedback mechanisms. This includes the direct comparison to monetary gains and losses, the use of uninformative, non-contingent feedback and no feedback as additional control conditions, as well as a differentiation between positive and negative feedback trials by using an event-related Scientific RepoRtS | (2020) 10:13030 | https://doi.org/10.1038/s41598-020-68300-y www.nature.com/scientificreports/ fMRI design with larger inter-trial intervals than the ones used in the current block design. Furthermore, examining larger groups of participants and including a manipulation of task difficulty might uncover additional brain regions besides the dPCC that are modulated by the informational value of feedback. Finally, we would like to argue that the interpretation of results on the neural correlates of feedback functions is not necessarily restricted to interactions within computer-controlled settings as used in most neuropsychological studies. The observation that feedback from technical systems is perceived and processed in a remarkably similar way to feedback from human interaction partners (see the CASA-Computers Are Social Actors-paradigm 67,68 ) suggests that such results offer insights into more general brain mechanisms of the various functions of feedback.

Methods
Participants. . They had to press a button with the right index finger in response to upward modulated FM tones and another button with their right middle finger indicating downward modulated FM tones. Immediately following the button press, system feedback was presented. Feedback in the blocks of stimuli differed with regard to informational content and modality as specified below. The experiment was divided into two functional runs that differed in the feedback's informational content. One half of all participants first received registering feedback and then evaluative feedback. The other half of the participants received evaluative feedback first. Registering feedback only indicated that the system has registered the participant's response without evaluating the correctness of their input. Evaluative feedback indicated if their input was correct or incorrect. Within each of the two consecutive runs, 24 blocks of stimuli were presented. Each block consisted of 16 FM tones presented with an inter-stimulus interval of 2 s. The blocks differed in feedback modality (eight blocks in each modality) and were presented in a pseudo-randomised order, alternating with resting blocks of 20 s. Depending on the block, feedback was presented either verbally (spoken feedback), in non-verbal auditory form (pure tones) or visually (symbols). Table 2 summarises the different types of feedback. Visual and auditory feedback were presented for 0.5 s. Verbal feedback had a similar duration ( 569 ms ± 26 ms).
During the entire experiment (with the exception of the periods of feedback presentation in the visual modality blocks), participants were asked to look at a white fixation cross on grey background.

Data acquisition. The measurements were carried out in a 3 Tesla scanner (MAGNETOM Trio, Siemens
Healthcare GmbH, Erlangen, Germany) equipped with an eight channel head coil. A 3D anatomical data set of the participant's brain (echo time (TE), 4.77 ms; repetition time (TR), 2500 ms; flip angle, 7 • ; matrix size, 256 × 256 voxels; field of view, 25.6 × 25.6 cm 2 ; 192 slices of 1 mm thickness) was obtained before the functional measurement. Additionally, a 2D anatomical data set was acquired before each of the two functional measurements using inversion-recovery echo-planar imaging (IR-EPI). The IR-EPI has the same geometric distortions as the functional measurement but a reversed contrast and thus allows a more precise coregistration of the functional data to the anatomical data. The two functional measurements each consisted of 634 functional volumes that were acquired in 21 min 8 s using an echo planar imaging (EPI) sequence (TE, 30 ms; TR, 2000 ms; flip angle, 80 • ; 3 mm isotropic resolution; 32 slices; 0.3 mm gaps).
The head of the participant was fixed with a cushion with attached ear muffs containing fMRI compatible headphones 71 . Additionally, the participants wore earplugs that, together with the headphones, reduced the scanner noise by 40-60 dB. The software Presentation (Neurobehavioral Systems Inc., Berkeley, CA) was used for stimulus presentation and the recording of behavioural responses. Before the experiment, the overall stimulus intensity was adjusted to a comfortable level for each participant and equally loud at both ears. Visual stimuli were presented by a video projector onto a back projection screen, which was visible inside the scanner via a mirror system. www.nature.com/scientificreports/ Data analysis. Behavioural data. Mean reaction times and error rates of each participant for each experimental condition were calculated from the behavioural log data of the fMRI experiments. These behavioural data were then further analysed with SPSS (IBM Corp., Armonk, NY) and examined for normal distribution by visual inspection of Q-Q plots, Shapiro-Wilk tests and evaluation of skewness and kurtosis. 2 × 3 factorial repeated-measures analyses of variances (ANOVAs) were conducted to examine significant differences in accuracy and reaction times depending on informational content (registering vs. evaluative feedback) and modality (visual vs. auditory vs. verbal feedback). To counteract violations of sphericity in effects including the 3-level factor modality, Greenhouse-Geisser (G-G) corrections were applied.
fMRI data. The functional data were analysed with the software BrainVoyager QX (Brain Innovation B.V., Maastricht, The Netherlands). A standard sequence of preprocessing steps, i.e. slice scan time correction, 3D-motion correction, linear trend removal, spatial smoothing with a Gaussian filter of 4 mm full width at half maximum, and temporal filtering with a high-pass of three cycles per scan was performed 10 . The functional data were inspected thoroughly for severe grey level fluctuations resulting from head motion. For that purpose, the automated head motion correction procedure, which resulted in estimated translation and rotation parameters for each spatial direction, was analysed 72 . Subjects with head movements that exceeded 2 • rotation or 3.0 mm translation in any direction in either of the two functional runs (evaluative feedback, registering feedback) were excluded from further analysis 70 . Finally, the functional data were co-registered with the 3D anatomical data by utilising the IR-EPI, and then transformed into Talairach-space 73 .
A random-effects analysis with a general linear model (GLM) including both z-transformed functional data sets of the 16 participants, was performed for all task conditions using the 2-gamma response function implemented in BrainVoyager QX. The model included one predictor for each of the six experimental conditions. To reduce signal artefacts from brain areas with low signal intensity, only those voxels were considered whose functional EPI signal had a grey level of at least 75.
To investigate the differential effects of evaluative versus registering feedback across modalities, the three modality-specific contrasts between evaluative feedback and registering feedback, and the three modality-specific contrasts between evaluative feedback and rest were combined in a conjunction analysis. To correct for multiple comparisons, a false discovery rate (FDR) of q = 0.01 (t > 3.40, p < 0.004) was applied. Volumes-of-interest (VOIs) were defined as all resulting clusters that comprised at least 54 mm 3 .

Data availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.