Musical performance is in essence a social conduct. It is a prime model for the ability to cooperate and work together toward a common goal. A deeper understanding of the neurophysiological mechanisms in musical performance might therefore be acquired by looking at it from the view point of task-sharing neurophysiological experiments (D'Ausilio et al., 2015). Such experiments have suggested that own actions and the observation of other's actions are functionally similarly represented in the brain (Knoblich and Jordan, 2003; Sebanz et al., 2003; van Schie et al., 2004; Bosbach et al., 2005; de Bruijn et al., 2012).

Ensemble musicians need to work together to achieve a satisfying performance. To reach this, they need to have common knowledge about how the whole piece should sound. During joint performance, musicians may then anticipate the actions of other performers in order to facilitate synchronization, and predict not only their own sound, but also the sound of the other ensemble members or the musical duet partner (Keller, 2008). The musical material of the partner is indeed of considerable importance since it constitutes an essential component for the common goal. Predictive mechanisms in auditory processing can be examined by using brain signals for prediction violations like the mismatch negativity (Fujioka et al., 2004; Kujala et al., 2007). The mismatch negativity (MMN) occurs in response to a deviation in a musical sequence that the subject listens to. An enlargement of the MMN to rhythmic and melodic deviations within a musical sequence is seen in musicians (Fujioka et al., 2004; Vuust et al., 2005), and can be observed already after 2 weeks of musical training in previously untrained subjects (Lappe et al., 2008, 2011). Source analyses revealed that the increased MMN component is generated in overlapping, but partially different brain networks depending on whether a melodic or rhythmic deviant occurred. Whereas melodic deviations are processed in a temporo-frontal network, rhythmic deviations activate neurons along the temporo-parietal pathway. (Lappe et al., 2013a,b, 2016). This is consistent with the dual pathway model of auditory processing (Belin and Zatorre, 2000; Arnott et al., 2004; Rauschecker and Scott, 2009; Bizley and Cohen, 2013).

During duet performance, a fundamental difference between the own and the partner material is that the former is played, and hence predicted based on one's own motor performance, while the latter is available only from listening to the auditory signal provided by the partner's performance. In previous studies on cortical plasticity after short-term musical training we found that only playing, but not mere listening, led to an increased mismatch negativity to the trained material (Lappe et al., 2008, 2011). The specific situation in ensemble play, however, might require predictive representations of the partner's material too, and should, thereby, lead also to plasticity effects for that material. In the present study, we were therefore interested in whether duet training induces enhancement in the musical MMN also for the partner material, and whether the sources of the mismatch detection were similar or different for deviations in one's own and the partner's material. In duet training, it is an important question whether only own errors in the own material are relevant for one's own learning, or whether the errors of the partner are equally important. Melodic and rhythmic errors of a musical duet partner have as much consequences as one's own errors on the musical performance, and the duet partners' mistakes presumably influence the learning process of the other musician. Consequently, when both errors are equally important and electrophysiologically similarly represented in the brain, the musically elicited MMN should increase after a joint musical training for a tone deviation within the own and the partner's musical part.

Error attribution, however, is also important in a duet performance (Loehr et al., 2013). Although a partner's error has similar consequences for the performance, pianists strategies for error avoidance and error management are presumably different when the error was committed by oneself as compared to the partner. A simultaneous training of two participants in a duet situation allows to differentiate between own and partner errors. We, therefore, also looked for a neural signature of the distinction between deviances in the own and the partner material.

We trained non-musician subjects in a piano-duet task and measured mismatch components elicited by musical deviations within the pianist's own and the partner's part before and after training. In a subsequent beamformer source analysis we investigated the brain areas contributing to the generation of the mismatch negativity and the distinction of deviations within subjects' own and partner material.

Participants increased their ability to play the piano exercise successfully during training. This is illustrated in Figure 2. Figure 2A shows the average number of correct intro sequences played per training session. Figure 2B illustrates the progress in tempo over the course of all training sessions.

From the MEG data, pronounced mismatch negativity components were noticeable in the RMS values after the tone deviation in the primo- and secundo parts before and after training. Figure 3 shows MMN RMS responses (deviant - standard) in the melody condition during the whole trial for the primo and secundo parts separately. Blue lines indicate responses before training, red lines indicate responses after training. In these figures, own and partner material are not yet separated. To study responses to own and partner deviations, we collected responses from both primo and secundo conditions and split these into own and partner material (Figure 4). Both, own and partner's material showed a clearly visible increase of the MMN after training in the melody as well as the rhythm condition. This observation was statistically supported in the melody condition by a within-subject ANOVA with factors own/partner and pretraining/posttraining, revealing a main effect of pretraining/posttraining [F_{(1, 15)} = 14.03; p = 0.04] and no interaction of own/partner X pretraining/posttraining. Similar results were found for the rhythm condition. A pretraining/posttraining main effect was also determined by a within-subject ANOVA with factors own/partner and pretraining/posttraining [F_{(1, 15)} = 7.19; p = 0.02]. No interaction effect was found between own/partner X pretraining/posttraining. We thus conclude that for both types of deviants the MMN increases after training, and, importantly, for one's own material as well as for the partner's material.

To determine the sources of the MMNs and a possible distinction between deviants in own and partners' parts, a beamformer analysis was conducted separately for the melodically and rhythmically elicited MMN and for the pianist's own and the partner's mismatch component. The melodically elicited MMN for pianists' own material revealed right sided neural activation in the anterior part of the auditory cortex (AC) and inferior frontal cortices (IFC) (Figure 5A, Table 1). Activation was also found in the precuneus and the medial frontal gyrus (Figures 5A,B) and, in addition, in the left cerebellum contralateral to the activation in the right auditory and inferior frontal areas (Figure 5C). Neural activation after a melodic deviation in the partner's material appeared less focussed, but also included activation of the AC in the left and right hemispheres (Figure 6A), the precuneus (Figure 6B), and the left cerebellum (Figure 6C). A direct comparison, however, of the pianist's own and partner's material did not reveal significant differences. We therefore conclude that a melodic deviation within a musical duet sequence elicits activation in the AC and IFC within one's own and the partner's material, representing a neural activation pattern that has been shown to also occur after a piano solo training (see Lappe et al., 2013a). In the piano duet situation of the present study, however, we find in addition neural activation in the precuneus, the medial frontal cortex as well as in the cerebellum in both players.

The rhythmically elicited MMN in the pianist's own material concurred with activations in the left inferior parietal area (Figure 7A), the posterior part of the left and right auditory cortices (Figure 7B) and the right cerebellum contralateral to the left inferior parietal lobule (Figure 7C). For the partner material, MMN activation was elicited in the left auditory cortex, the cingulum and the left superior frontal and medial frontal cortex (Figures 8A,B). A direct comparison between own and partner material showed also no differences. We therefore conclude that a rhythmic deviation within a musical duet sequence elicits neural activation in the left and right AC and the left IPL within one's own and the partner's material, which is an activation pattern that has previously been shown to occur following a rhythmic deviation after a piano solo training (Lappe et al., 2013b). In the piano duet situation of this study, however, we find, in addition, neural activation in the frontal and medial frontal cortices as well as in the cerebellum in both players.

In the present study, non-musician subjects learned to play a musical exercise within a piano duet. Before and after training, we measured the musically elicited mismatch negativity that was generated by listening to deviations within the pianist's own or the partner's material. The MMN increased significantly after training for both parts. These plastic changes were observed both in the melodically and the rhythmically elicited MMN. This finding extends previous results showing training-induced plasticity effects only in musical material that was motorically trained but not after pure auditory training (Lappe et al., 2008, 2011; Maidhof et al., 2010). Specifically, in the study by Lappe et al. (2008) the MMN was larger after sensorimotor training compared to an auditory control group in which subjects listened to the same training material but did not play. The study by Maidhof et al. (2010) showed that the error related negativity (ERN) is larger when built up during musical performance compared to mere listening. The results of both studies suggest that expectancy is stronger when established during motor performance. The results of the present study support the hypothesis that in a duet situation, the partner's part, although only heard and not motorically learned, has a stronger significance compared to a mere listening situation. Duet partners need to work together to achieve a common, satisfying performance. The partner's part is therefore highly relevant for one's own progress and for the common goal, which therefore presumably leads to neural representations of the partner's musical part, too. Moreover, mistakes in the partner's part are highly important and discrimination accuracy, therefore, increases significantly also for the partner's musical material (Loehr et al., 2013).

Such a training-induced increase in the MMN to rhythmic or melodic deviation within a musical context points to an audio-motor coupling after sensorimotor training (Baumann et al., 2005; Bangert et al., 2006; Halwani et al., 2011; Jäncke, 2012; Mathias et al., 2015) leading to a better musical discrimination accuracy on the one hand and a better prediction ability about upcoming musical events on the other. It has been suggested that musical training establishes an internal forward model linking a musical tone with a specific motor movement resulting in strong associations between actions and their auditory consequences (Lee and Noppeney, 2011). Such an internal forward model would then predict the consequences of these actions by using efference copies of the motor command (Wolpert et al., 1995) helping the neural network to better detect auditory prediction violations, as manifested by the musical MMN. It has furthermore been suggested that, in a musical duet situation, each musician, since he or she learns to predict what the musical partner is doing, simulates the partner's musical part and establishes thereby an internal forward model for the partners material, too (Novembre and Keller, 2014; Hadley et al., 2015). Our results confirm this hypothesis. Duet pianists might have learned to form expectations not only by associating their own actions, but also by associating their partner's actions with specific auditory outcomes.

The mismatch negativity components generated for own and partner material were based on physically identical stimuli, thereby enabling a direct comparability of the mismatch enhancement after training. The associated motor command during training and the internal model of the outcome, however, were different since the tones had to be played at different time points in the training material.

Errors are important for learning. Brain activation related to errors in piano playing can be seen in components of electroencephalographic (EEG) data even before the onset of motor activity (Maidhof et al., 2009; Ruiz et al., 2009). The error-related negativity (ERN) is stronger during performance than during listening (Maidhof et al., 2010). Likewise, the mismatch negativity becomes stronger after training (Lappe et al., 2008, 2011). Larger error related ERF components following an error during or after a musical performance are presumably an indication that these errors are perceived by the musician as being highly relevant and important. In duet training, it is an essential question whether only own errors in the own material are relevant for one's own learning or whether the errors of the partner are equally important. Our results suggest that the latter is the case. A common performance is only possible when both players are able to accomplish their parts, and own and partner mistakes have equal consequences for the common goal. The increase of the MMN component for melodic and rhythmic deviations in the partner's part, as it was shown in our study, is therefore evident and plausible.

However, to learn from errors and to improve duet performance, players have to distinguish own errors from errors of the partner, attributing errors in relation to agency. Therefore, although the error related and the mismatch negativity components share similar neural representations, the requirement to attribute an error to oneself or to the partner has to be reflected somewhere in the neural network. Previous fMRI and ERP studies have mainly shown similar responses to one's own and to other's errors (de Brujin and Rhein, 2012; de Bruijn et al., 2012; Loehr et al., 2013). Loehr et al., for example, found similar feedback-related negativities for altered auditory stimuli during a piano duet performance no matter whether they occurred within the own or the partner's part. However, a larger P300 was seen within the own compared to the partner's material, suggesting that a differentiation of own and partners' action is processed at later times. Our beamformer results of own and partners' mismatch responses indicate also that the same networks are involved for own and partners' tone deviations since a direct subtraction of own and partner MMNs did not yield significant results. We therefore concentrate on those activation patterns that are similar for own and partner material. We find comparable neural activation for own and partner material in the melody condition, revealing neural activity in the auditory cortex in combination with inferior frontal cortex. This activation pattern has been shown in previous auditory deviance detection studies, especially after short- and long- term musical training (Lappe et al., 2013a,b, 2016) indicating that melodic deviations are processed in temporal-frontal networks. In addition, however, strong activation was found in the present duet study in the precuneus and the right superior medial frontal cortex. These brain areas have been implied in previous studies in the distinction between own and partners' actions (Frith and Frith, 1999; Amodio and Frith, 2006; Cavanna and Trimble, 2006; Isoda and Noritake, 2013). Hence, our data show that a neural signature of the necessity to differentiate between own and partner errors is already apparent at the early timing of the mismatch negativity component.

A further interesting point concerns the involvement of the cerebellum. The cerebellum is considered difficult to detect with MEG, since it is positioned rather distant from the sensors. Several recent MEG studies, however, have shown activation in cerebellar networks (Krause et al., 2010; Wibral et al., 2010; Fujioka et al., 2014; Lu et al., 2014). In the present study, cerebellar activation is consistently observed in almost all conditions and is clearly contralateral to the cortical activity, which is consistent with the opposing organization between cerebellar cortex and neocortex. Therefore, we are confident that the cerebellar activation in our data is of relevance. In fact, the cerebellum is considered an important part of the forward model of many motor behaviors (Wolpert et al., 1998). We therefore suggest that its involvement in the present study was built during the sensorimotor training process. However, since the mismatch measurements themselves do not involve motor behavior, our results suggest that the cerebellum contributes to prediction not only of motor actions but also of sensory sequences, consistent with recent findings from sensory and motor timing studies (Kotz et al., 2014).

This is also seen in the activation pattern after a rhythmic deviation in the own material. It revealed an involvement of the left and right auditory cortices, the left inferior parietal lobule, and the right cerebellum, contralateral to the left inferior parietal activity. This finding is again consistent with the cerebellum being part of the forward model built during the sensorimotor training. The fact that we found activation for the melody deviation in the left cerebellar hemisphere and for the rhythmic deviation in the right cerebellar hemisphere is especially interesting since it supports the general assumption that melodic sequences are processed more strongly in the right neocortical hemisphere whereas rhythm is stronger processed in the left neocortical hemisphere (Zatorre and Belin, 2001).

An involvement of the motor system in predicting musical sequences in the mismatch paradigm is sometimes, particularly in musicians, also shown by activation of cortical motor or premotor areas (Lappe et al., 2016). This was not the case in the present study. Perhaps a longer training is necessary to observe motor activation during the mismatch response using MEG and beamformer analysis. It is also possible that motor activation becomes visible in later components, such as the N200 (Mathias et al., 2015) that we could not analyze in our data set.

In summary, our results provide a view on the neural underpinnings of musical ensemble performance that suggests a common representation of the musical piece, acquired during joined training and established from predictive processes of sensorimotor performance together with an involvement of structures that differentiate the sources of any deviations.

CL designed and conducted the experiment, analyzed the data and wrote the manuscript. SB performed the training. ML analyzed the data and wrote the manuscript. CP read and revised manuscript.

The present work has been funded by the German Research Foundation DFG LA 3080/1-1.