The present functional magnetic resonance imaging study investigated the instruction-based learning of novel arbitrary stimulus–response mappings in order to understand the brain mechanisms that enable successful behavioral rule implementation in the absence of trial-and-error learning. We developed a novel task design that allowed the examination of rapidly evolving brain activation dynamics starting from an explicit instruction phase and further across a short behavioral practice phase. As a first key result, the study revealed that different sets of brain regions displayed either decreasing or increasing activation profiles already across the first few practice trials, suggesting an impressively rapid redistribution of labor throughout the brain. Furthermore, behavioral performance improvement across practice was tightly coupled with brain activation during the practice phase (caudate nucleus), the instruction phase (lateral midprefrontal cortex), or both (lateral premotor cortex bordering prefrontal cortex). Together, the present results provide first important insights into the brain systems involved in the rapid transfer of control from initially abstract rule representations induced by explicit instructions toward pragmatic representations enabling the fluent behavioral implementation.
One of the outstanding human abilities is to instantaneously implement novel, completely arbitrary behavioral instructions. This type of instruction-based learning stands in stark contrast to the rather time-consuming learning by trial and error, which is required when no clear instruction is given or when the complexity of the instruction is beyond working memory capacity (Petrides 1997). Experimental psychologists naturally exploit this remarkable learning ability when they instruct human participants for the most arbitrary tasks. Indeed, 2 observations can be commonly made in experimental psychology laboratories. First, an attentive participant has typically no difficulty in implementing the task instructions right from the first trial. Second, it takes only a few more practice trials before the participant seems to start feeling comfortable with the task. The present study sought to characterize the neurocognitive processes that are supposed to be taking place in the initial phase of instruction learning. Our main focus was on the first few practice trials that required the actual behavioral implementation of newly instructed rules, but we also examined the relevance of the initial rule encoding during the preceding instruction phase. We were particularly interested in this early phase of practice led by the theoretical assumption that the behavioral expression of a newly instructed task should crucially rely on the successful transfer of control from of a highly abstract, possibly verbally coded task representation formed during the instruction phase toward a more “pragmatic” task representation required for actual motor implementation during behavioral practice trials (Nakayama et al. 2008). In fact, it might be this transfer of control from abstract stimulus–response (S–R) rules toward a pragmatic code that does not work properly in patients with prefrontal damage who are reportedly unable to behaviorally implement a novel instruction while being able to repeat the instruction verbally (Luria 1973).
We created an experimental task in which participants were required to sequentially implement 20 different behavioral instructions each involving 4 unique stimuli mapped onto 2 manual responses (Fig. 1A). Following each novel instruction, there were 32 practice trials, during which each S–R association was repeated 8 times in a randomly intermixed fashion. This task design enabled us to directly examine the brain activation dynamics associated with rapid incremental adaptation processes following the explicit instruction of novel S–R associations (see Materials and Methods). By contrast, previous neuroscientific research has almost entirely focused on the acquisition of novel S–R mappings based on trial-and-error learning, including neuroimaging studies on human subjects (Deiber et al. 1997; Toni et al. 2001; Eliassen et al. 2003; Boettiger and D'Esposito 2005; Brovelli et al. 2008) and studies based on intracranial neuronal recordings or lesions in monkeys (Asaad et al. 1998; Wang et al. 2000; Bussey et al. 2001; Wirth et al. 2003; Brasted and Wise 2004; Pasupathy and Miller 2005). Such studies suggest important roles of premotor, lateral prefrontal, striatal, and mediotemporal areas in the acquisition of arbitrary S–R mappings (Murray et al. 2000; Passingham et al. 2000; Wise and Murray 2000; Petrides 2005a, 2008; Suzuki 2007).
Although learning by trial-and-error seems profoundly different from instruction-based learning, neuropsychological studies based on human patients with lesions across the lateral prefrontal cortex (LPFC) tentatively suggest at least a partial overlap in the cerebral implementation of these different types of learning (Petrides 1990, 1997). Thus, albeit far from facing a clear picture with regard to the commonalities and differences of both types of learning, some conclusions from the trial-and-error learning literature appear sufficiently general to warrant predictions regarding the functional neuroanatomy of instruction-based learning. In particular, results from lesion studies in monkeys suggest that the midventrolateral prefrontal cortex (mid-VLPFC) might be necessary for the learning of novel S–R mappings but not for the implementation of well-learned associations (Murray et al. 2000; Wang et al. 2000; Bussey et al. 2001). In contrast, an area at the border of the anterior premotor cortex and the adjacent posterior PFC (from hereon termed “lateral premotor cortex–prefrontal cortex” or “lateral PMC–PFC”) needs to be intact during both rule learning and continued future rule implementation (Petrides 2005). The lateral PMC–PFC region has been referred to as “periarcuate region” in monkeys, but because the present study investigated humans, we will use the terminology for human neuroanatomy when referring to this region. Such findings corroborate the interpretation that mid-VLPFC might be relevant for rule extraction during trial-and-error learning, whereas the lateral PMC–PFC might be relevant for representing the learned S–R associations. Yet, based on trial-and-error learning procedures, it seems difficult, if not impossible, to discern the distinct contributions of multiple, temporally intermingled subprocesses, including 1) the application of generalized strategies for rule extraction, 2) the formation of abstract (i.e., verbal) rule representations, and 3) the formation of pragmatic rule representations. For example, it is not clear whether the involvement of the mid-VLPFC during learning reflects the application of generalized rule extraction strategies or rather the buffering of incrementally extracted abstract S–R rules within working memory. Similarly, abstract rule formation and pragmatic rule implementation are difficult to discern as these 2 processes are, at least partially, overlapping in time.
In instruction-based learning, some of the above-mentioned subprocesses are seemingly irrelevant, whereas others appear to be much more clearly defined. First of all, rule extraction is, by definition, not relevant in instruction-based learning—at least as long as S–R mappings are sufficiently simple to be kept in working memory. Second, the formation of abstract rule representations should be completed by the end of the instruction period—as should be evidenced by low error rates right from the first practice trial. By contrast, the other relevant aspect of instruction-based learning, that is, the recoding of abstract task rules into pragmatic S–R representations, was assumed to start from the first practice trial, followed by a consolidation period during which the pragmatic representation becomes further strengthened and thus less and less dependent on support from the abstract rule representation. Accordingly, we hypothesized that the transfer of action control from an abstract to a pragmatic task representation should be indicated by a rapid redistribution of labor among brain regions across the first few practice trials following the instruction phase. Specifically, we expected a strong initial engagement of brain areas that support the retrieval and implementation of novel abstract task rules, including the mid-VLPFC (Wang et al. 2000; Bussey et al. 2001), possibly in cooperation with mediotemporal brain structures (Yanike et al. 2009). Such brain regions were expected to become less important with increasing stability of the newly formed pragmatic representation, reflected by an increasing engagement of other brain regions including the lateral PMC–PFC possibly in cooperation with striatal subregions (Brasted and Wise 2004).
Most importantly, we reasoned that the true significance of such activation dynamics would eventually need to stand the test against its implications for actual behavioral performance—even more so, as our theoretical emphasis was put on the pragmatics of explicitly instructed rules. Therefore, we first examined the relationship between practice-related brain activation dynamics and behavioral performance. The rationale was that a more efficient transfer of control toward a pragmatic task representation would lead to a stronger performance improvement across the course of practice. Thus, practice-related activation dynamics in brain regions that are specifically involved in the hypothesized transfer of action control from an abstract to a pragmatic task representation were expected to covary with practice-related performance improvement. Second, we exploited one of the strengths of the cognitive neuroscience approach, that is, to be able to investigate the neural correlates of cognitive processes in the absence of overt behavior. Specifically, we determined the correlation between brain activation during the instruction phase and the subsequent practice-related behavioral performance improvement. This analysis allows the identification of brain regions that are essential for the encoding of task rules in a way that is suited to successfully guide future behavioral rule implementation.
Materials and Methods
Sixteen right-handed human participants with no evidence of neurological or psychiatric compromise took part in this study (mean age = 26 years; age range: 21–30 years; 11 females and 5 males). The experimental protocol was approved by the Ethics Committee of the Technische Universität Dresden. All participants gave written informed consent prior to taking part in the experiment and were paid € 8 per hour for their participation.
We devised an experimental procedure that required participants to sequentially implement 20 different, explicitly instructed arbitrary S–R mappings. Each mapping involved 4 unique abstract shapes, 2 of which required a left index finger button press and the other 2 required a right index finger button press (Fig. 1, upper panel). Each mapping was instructed visually on the computer screen for 10 s. During this instruction phase, 2 stimuli were displayed left of the center (indicating that a left response was demanded), and the other 2 stimuli were displayed right of the center (indicating that a right response was demanded). The instruction was followed by 32 practice trials, each trial comprising one stimulus that was displayed until the participants’ response or until the maximum response time window of 1500 ms had elapsed. Error trials (or misses due to time-out) were repeated in the following trial, and both trials were excluded from the analysis. Participants received feedback about the correctness of a response that was displayed for 500 ms immediately after both correct responses and erroneous responses or after time-out. Because there were only 2 possible responses available, an erroneous response indicated unambiguously that the alternative response would have been the correct choice. Unique random sequences of the 4 stimuli were generated for each mapping and for each participant such that each individual stimulus occurred 8 times during practice, not including the additional presentations after erroneous responses. The intertrial interval varied randomly between 0.8 and 3.5 s. The total experiment duration was approximately 35 min, depending on response speed and the number of erroneous responses.
Whole-brain images were acquired on a Siemens 3-T whole-body Trio System (Erlangen, Germany) with a 16 channel circularly polarized head coil. Headphones (NordicNeuroLab, Bergen, Norway) and earplugs dampened scanner noise. Both structural and functional images were acquired for each participant. High-resolution structural images (1.0 × 1.0 × 1.0 mm) were acquired using an magnetization-prepared rapid gradient echo T1-weighted sequence (time repetition [TR] = 1900 ms, time echo [TE] = 2.26 ms, time to inversion = 900 ms, flip = 9°). Functional images were acquired using a gradient echo-planar sequence (TR = 2000 ms, TE = 30 ms, flip = 80°, interleaved slice acquisition, slice gap = 0). Each volume contained twenty-six 5.0-mm thick slices (in-plane resolution, 4.0 × 4.0 mm). The experiment was controlled by Presentation 12.0 software (Neurobehavioral Systems, San Francisco, CA) running on a Windows-XP PC. Stimuli were projected to participants via Visuastim digital goggles (Resonance Technology, Inc.; Northridge, CA) simulating a viewing distance of 100 cm. A fiber optic, light-sensitive key press was used to record participants’ behavioral responses.
The empirical data set were analyzed with SPM5 based on MATLAB 7.1. Preprocessing included slice-time correction, rigid body movement correction (3 translation and 3 rotation parameters), normalization of the functional images by directly registering the mean functional image to the standard Montreal Neurological Institute echo-planar-imaging template image provided by SPM5 (the resulting interpolated spatial resolution was 4 × 4 × 4 mm), and smoothing of the functional images (Gaussian Kernel, full width half maximum = 8 mm).
The preprocessed imaging data were analyzed using the general linear model (GLM) approach as implemented in the SPM5 software package. Model regressors were created by convolving neural input functions for the different event types with the assumed canonical hemodynamic response function used by SPM5. The analysis aimed at estimating blood oxygen level–dependent (BOLD) activation associated with 1) the 10-s instruction phase and 2) the subsequent practice phase.
Instruction-related activation was captured by 2 different types of regressors both synchronized to the onset of the instruction phase, which were estimated within 2 separate GLMs. One regressor was constructed to capture BOLD activation associated with transient neural activity at the start of the instruction phase (duration parameter = 0). The other regressor was constructed to capture BOLD activation associated with sustained neural activity throughout the instruction phase (duration parameter = 10). Note that implementing an “integrated” GLM solution, in which both instruction-related regressors were estimated within the same model, yielded highly similar results (not reported).
The successful extraction of practice-related BOLD activation relied on an important property of the task design. Specifically, the presentation order of the 4 instructed stimuli during the practice phase was randomly intermixed for each of the 20 S–R mappings. Thus, it was possible to transform the series of 32 successive practice trials (represented on the x-axis in Fig. 1, lower panel) into an intermixed sequence of 8 stimulus repetition levels for each of the 4 stimuli (represented on the y-axis in Fig. 1, lower panel). Thus, instead of trying the impossible, that is, to estimate the highly correlated BOLD responses for each of the 32 successive practice trials, it was possible to analyze practice-related dynamics in terms of the incremental number of repetitions of the 4 S–R associations yielding a highly reduced correlation among BOLD responses for the 8 stimulus repetition levels (Eliassen et al. 2003; Brovelli et al. 2008). Given 20 different S–R mappings, for each of the 8 stimulus repetitions, there were 80 (20 × 4) correct trials available for analysis. We included one additional model regressor which captured activation elicited by error and posterror trials.
The activation dynamics across the 8 stimulus repetition levels was modeled differently within 3 separate GLMs, whereas the model regressors for the instruction phase and the error and posterror trials were always the same. In GLM-1 and GLM-2, stimulus repetition was modeled via parametric designs, whereas GLM-3 realized a factorial design. In GLM-1, we included a single regressor for all practice trials (i.e., 20 × 32 event occurrences) and captured practice-related changes of activation by adding a parametric variable that varied with stimulus repetition level, assuming a linear relationship between stimulus repetition level and BOLD activation change (i.e., parameter values ranged between 1 and 8 according to stimulus repetition level). GLM-2 was identical to GLM-1 except that the relationship between stimulus repetition level and BOLD activation change was assumed to follow an exponential decay function (i.e., parameter values from the linear model were transformed via an exponential decay function). In both GLMs, the presence of practice-related BOLD activation change was evaluated by performing t-tests for the respective stimulus repetition parameter. In contrast, the factorial design of GLM-3 comprised 8 regressors, one for each stimulus repetition level. Within this model, the presence of practice-related activation change was evaluated based on an F statistic testing for significant differences among the 8 stimulus repetition levels. The advantage of the parametric GLMs is that practice-related activation dynamics that resemble either a linear function or an exponential decay function can be captured more reliably due to fewer degrees of freedom. The clear disadvantage is that detailed information about activation at each stimulus repetition level gets lost. Thus, conclusions can only be drawn regarding the direction of practice-related activation change (i.e., increasing or decreasing activation with increasing number of stimulus repetitions). For instance, it is impossible to determine whether a voxel that exhibits decreasing activation is still significantly activated above baseline at the end of practice (i.e., at stimulus repetition 8). Another potential disadvantage of the parametric model as compared with the factorial model is that other relationships between stimulus repetition and BOLD signal change (e.g., U shaped) might be overlooked. Yet, it seems that, at least in this study, the relevant practice-related activation dynamics were fully captured by the 2 parametric models as the factorial model did not identify voxels exhibiting significant practice-related activation that were not also identified by either of the parametric models. Notably, also the 2 parametric models yielded quite similar results.
We used GLM-1 and GLM-2 to identify voxels that showed significant practice-related increase or decrease of BOLD activation (parametric contrast p(t) < 0.001; uncorrected). These voxels were further assessed by the factorial GLM-3 for 2 reasons. First, GLM-3 provides BOLD activation estimates for each of the 8 stimulus repetition levels, thus allowing for a detailed description of the actual practice-related activation dynamics. Second, we used the BOLD response estimates for stimulus repetitions 1 and 8 in several inclusive and exclusive masking procedures to subdivide the previously identified practice-related voxels into 4 mutually exclusive categories. Those included 1) voxels exhibiting practice-related activation decrease terminating significantly above baseline, that is, a significant negative slope for either parametric contrast [p(t)< 0.001] and a significant positive amplitude for the factorial contrast at stimulus repetition 8 (p(t) < 0.05); 2) voxels exhibiting practice-related activation decrease terminating at zero activation, that is, a significant negative slope for either parametric contrast (p(t)< 0.001) and no significant positive activation for the factorial contrast at stimulus repetition 8 [p(t)> 0.05]; 3) voxels exhibiting practice-related activation increase starting at or above baseline, that is, a significant positive slope for either parametric contrast [p(t) < 0.001] and no significant negative activation for the factorial contrast at stimulus repetition 1 [p(t)> 0.05]; and 4) voxels exhibiting practice-related activation increase below baseline, that is, a significant positive slope for either parametric contrast [p(t)< 0.001], a significant negative activation for the factorial contrast at stimulus repetition 1 [p(t)< 0.05], and no significant positive activation for the factorial contrast at stimulus repetition 8 [p(t)> 0.05]. Finally, we included a fifth category for voxels exhibiting significant activation during the practice phase at a constant level, that is, a significant effect for the overall “trial” regressor estimate in either GLM-1 or GLM-2 [p(t)< 0.001] and no significant effects of stimulus repetition [p(t)> 0.05] for either parametric contrast. Finally, we excluded voxels from this fifth category that were located within an 8-mm radius of voxels showing either a significant activation increase or a significant activation decrease. This was done to avoid erroneous classification due to the mutual cancellation of increasing and decreasing activation profiles from closely spaced foci, which are likely to merge due to spatial smoothing and averaging across interindividually variable brain morphologies.
To control for multiple comparisons, all activation maps were additionally assessed via the cluster-level criterion based on the Gaussian-random field theory implemented in SPM5. Given the particular smoothness properties of the present data set this means that an activation cluster (with an initial uncorrected voxel-level threshold set to P < 0.001) needed to comprise at least 26 contiguously activated voxels to reach a cluster-corrected significance level of P < 0.05. For a few cases during the analysis of the relationship between performance improvement and BOLD activation (see next section), we relaxed these demands in order not to miss potentially relevant information. Importantly, each case is explicitly identified.
Performance Improvement × BOLD Activation
Analogously to the analysis of the brain activation data, we examined practice-related changes of behavioral performance as a function of stimulus repetition level with regard to both the mean response times and the percentage of errors. Furthermore, and most importantly, we examined the relationship between practice-related behavioral performance improvement and BOLD activation. To determine the covariance between behavior and brain activation, the behavioral performance improvement based on response times for each participant was entered as a covariate into the statistical parametric mapping (SPM) group-level analysis of 1) practice-related BOLD activation dynamics (based on the parametric stimulus repetition contrasts from GLM-1 and GLM-2) and 2) instruction-related BOLD activation. The improvement of behavioral performance was expressed by a single parameter obtained by fitting an exponential decay function across the mean response times for each stimulus repetition level, with higher positive decay values indicating a stronger performance improvement. To make sure that the SPM-based regression analysis did not produce significant correlations due to possible outliers, we additionally computed the Spearman rank correlation between performance improvement and brain activation from the peak activation voxel in each region identified via the original SPM regression analysis (weighing reduced statistical power against insensitivity to outliers). The scatter plots depicted in Figures 5 and 6 are based on the ranked values for behavior and brain activation that were used for the Spearman correlations.
In a first analysis step, we determined practice-related changes in behavioral performance. The results are summarized in Figure 2 depicting mean response times and error rates as a function of stimulus repetition. Clearly, both curves are characterized by high learning rates, leading to asymptotic performance levels around stimulus repetition 4. According to a repeated measures analysis of variance, the general performance improvement with increasing stimulus repetition was significant for both response time (RT) (F3,49 = 20.9; P < 0.001, Greenhouse–Geisser-corrected) and error rate (F5,70 = 4.2; P = 0.003, Greenhouse-Geisser-corrected). With regard to RT, the pairwise contrasts for adjacent stimulus repetitions were significant for stimulus repetition 2 versus 1 (F1,15 = 25.3; P < 0.001) and for stimulus repetition 3 versus 2 (F1,15 = 15.8; P < 0.001). No other such pairwise contrasts were significant (neither for RT nor for error rate). Thus, the performance data suggest that some rapid adaptation processes might have been completed already halfway through the practice phase. Moreover, the error rate was impressively low right from the first stimulus repetition (mean 6%; ranging between 0% and 15% across participants) and decreased by only 3% until the end of practice. These observations suggest that the instructions were well memorized, and thus, trial-and-error learning seems unlikely to have contributed substantially to the observed performance improvement.
The imaging data corroborate the conclusion that rapid adaptation processes have occurred across the first few stimulus repetitions. We will first describe the general pattern of practice-related BOLD activation changes (i.e., different types of activation increase and decrease). Thereafter, we will turn to the key analysis investigating the relationship between practice-related performance improvement and brain activation during 1) the practice phase and 2) the instruction phase.
We identified 4 types of practice-related BOLD signal change, including 2 types of BOLD signal decrease (Supplementary Tables S1 and S2) and 2 types of BOLD signal increase (Supplementary Tables S3 and S4). Furthermore, we identified several brain regions that exhibited significant BOLD activation during the practice phase at a constant level, that is, irrespective of the stimulus repetition (Supplementary Table S5; see for a representative brain region, Fig. 4B).
Practice-Related Activation Decrease
Brain areas exhibiting a BOLD signal decrease included the LPFC along the inferior frontal sulcus, caudally transitioning into the inferior frontal junction (IFJ) area (Brass et al. 2005; Derrfuss et al. 2005). The same overall decreasing pattern was also observed along the intraparietal sulcus (IPS) and in the anterior insular cortex. Interestingly, we observed stark differences between voxels with regard to their asymptotic levels of activation. Specifically, some voxels reached asymptotic activation well above baseline level, whereas other voxels reached activation levels close to baseline (i.e., close to zero). To further objectify this observation, we created 2 separate whole-brain activation maps based on the original map for practice-related BOLD signal decrease (see Materials and Methods). Because we were particularly struck by the observation of near-baseline asymptotic activation levels (especially in the LPFC), we chose a very lenient criterion of P < 0.05 (uncorrected) to determine the activation level for the final stimulus repetition 8 in order to avoid misclassifying weak asymptotic activation as nonexistent activation. Thus, only voxels exhibiting close-to-zero activation at stimulus repetition 8 would become classified as nonsignificantly activated above baseline at the end of practice (for a representative brain regions, see Fig. 3C). Conversely, voxels exhibiting only weak activation at stimulus repetition 8 would already be classified as significantly activated above baseline at the end of practice (for a representative brain regions, see Fig. 3B).
This additional subclassification into 2 types of BOLD signal decrease revealed that activation in most of the LPFC declined to baseline level at the end of practice. The only exceptions were the more caudolateral part of the IFJ (bilaterally) and the right mIFS. These regions exhibited significant, though numerically small, above-baseline activation even at the end of practice. Similar to the predominant activation pattern in the LPFC, a BOLD signal decrease down to baseline level was also observed for the lower bank of the IPS. By contrast, the BOLD signal recorded from the upper bank of the IPS remained well above baseline until the end of practice. Notably, as exemplified by the 2 representative brain regions depicted in Figure 3B, significant asymptotic above-baseline activation was consistently observed on a much higher general level of activation in parietal regions as compared with frontal regions.
Practice-Related Activation Increase
Two completely different types of practice-related BOLD signal increase were observed, including 1) above-baseline BOLD signal increase with increasing stimulus repetition (for exemplary brain regions, see Fig. 3D) and 2) below-baseline BOLD signal increase with increasing stimulus repetition (for an exemplary brain region, Fig. 4A).
We identified several regions that exhibited above-baseline BOLD signal increase, including the precentral/postcentral gyri dorsomedially transitioning into the supplementary motor area (SMA) and ventrally transitioning into the rolandic operculum, and the adjacent middle insular cortex and superior temporal gyrus. Additionally, a significant practice-related BOLD signal increase was found in the striatum. As mentioned above, we also observed brain areas exhibiting practice-related BOLD signal increase starting from (and remaining) below baseline (Supplementary Table S4). These regions by and large overlap with the resting state or default-mode network described previously (Raichle et al. 2001; Miall and Robertson 2006). Accordingly, it seems likely that the practice-related activation increase in these regions is rather unspecifically related to a gradual return to the resting state. In other words, the system becomes increasingly less absorbed by the experimental task, thus freeing increasingly more resources for resting state processes.
Relationship between Practice-Related Activation and Behavioral Performance Improvement
So far, our results have strikingly revealed the presence of rapid adaptation processes following the explicit instruction of novel S–R mappings as evidenced by the practice-related dynamics of both behavior and brain activation. The next, most crucial, analysis step determined the extent to which practice-related changes in BOLD activation were directly associated with practice-related performance improvement. This analysis was meant to provide much more direct evidence for the potential involvement of certain brain regions in the formation of pragmatic rule representations (as actual behavior should be closely controlled by this representation). We found 3 brain regions that exhibited significant negative correlations between practice-related changes in performance and practice-related changes in brain activation (Fig. 5; Table 1), including a region at the border of the right lateral PMC and PFC (at a lower significance threshold of P < 0.01, uncorrected, also in the left hemisphere at MNI coordinates: −56, −4, 36), the head of the right caudate nucleus (at a lower significance threshold of P < 0.01, uncorrected, also in the left hemisphere at MNI coordinates: −8, 20, 8), and a region in the right posterior superior parietal lobule (pSPL). Note that only the right PMC activation conforms to the whole-brain cluster-level–corrected statistical criterion equivalent to at least 26 contiguously activated voxels. By contrast, the right pSPL reached only voxel-level significance at P < 0.001 (uncorrected). The right caudate nucleus failed the cluster-level-corrected threshold on the whole-brain level but reached cluster-level-corrected significance after small volume correction selectively including voxels that exhibited a general practice-related activation increase above baseline.
|Brain region||MNI coordinates||Statistics|
|x||y||z||BOLD (practice) × performance improvement||BOLD (instruction) × performance improvement|
|R. caudate nucleus||12||20||4||15||−5.00***b||−0.68**||−0.54|
|Brain region||MNI coordinates||Statistics|
|x||y||z||BOLD (practice) × performance improvement||BOLD (instruction) × performance improvement|
|R. caudate nucleus||12||20||4||15||−5.00***b||−0.68**||−0.54|
Note: Asterisks denote significant t-values and Rho values: *P < 0.05, **P < 0.01, ***P < 0.001, uncorrected at the voxel level. Abbreviations: R, right; a, anterior; p, posterior.
Significant activations with P < 0.05, corrected at the cluster level (i.e., at least 26 contiguous voxels).
Significant activations with P < 0.05, corrected at the cluster level after small volume correction selectively including voxels that exhibited a practice-related activation increase. Rho denotes Spearmans correlation coefficient. All reported correlations for BOLD (instruction) × performance improvement are based on the BOLD estimate for transient instruction-related activation.
Importantly, the lateral PMC–PFC region corresponds roughly to the periarcuate region in monkey brains described by Petrides (2005), a region that we specifically predicted to be involved in the formation of pragmatic rule representations. This relatively large right-hemispheric activation cluster transitioned posteriorly into a region that showed practice-related activation increase and anteriorly into a region that exhibited practice-related activation decrease, partially overlapping with the IFJ area defined above. Thus, for these 2 PMC-PFC subregions (from hereon termed as PMC-p and PMC-a) the actual performance-dependent activation dynamics underlying the negative correlations between performance and BOLD activation should be quite different. Specifically, a negative correlation between performance improvement and activation increase as observed in the PMC-p (and in the caudate nucleus) indicates weaker activation increase (i.e., BOLD parameter estimate less positive) with higher performance improvement. By contrast, a negative correlation between performance improvement and activation decrease as observed in the PMC-a (and in the right pSPL) indicates stronger activation decrease (i.e., BOLD parameter estimate more negative) with higher performance improvement. These 2 contrasting patterns of performance-dependent activation dynamics are visualized in Figure 6. Specifically, practice-related BOLD activation dynamics are shown as a function of practice-related performance improvement after implementing a median split separately for each of the 2 PMC subregions. Note that this median split was intended for visualization purposes only. For the statistical assessment of the relationship between performance and brain activation, the results from the regression analyses are preferable (MacCallum et al. 2002). The median split visualization clearly demonstrates the expected differential activation patterns for the anterior and posterior PMC subregions (Fig. 6A,B). BOLD activation decrease in the PMC-a was stronger for high as compared with low performance improvement. By contrast, the PMC-p showed the opposite pattern, that is, stronger activation increase for low as compared with high performance improvement.
It seems tempting to interpret these contrasting patterns of performance-dependent activation dynamics in terms of a direct interaction between both subregions assuming that PMC-a provides abstract information about the instructed rules as long as a stable pragmatic code has not yet been established within PMC-p. Thus, when performance improvement is low, that is, when pragmatic code formation is supposed to be rather inefficient, the PMC-a needs to provide abstract rule information for a longer period of time throughout the practice phase (little PMC-a activation decrease), paralleled by a more pronounced increase of PMC-p recruitment, reflecting the ongoing formation, and strengthening of the pragmatic representation. Conversely, when performance improvement is high, that is, when pragmatic code formation is supposed to be highly efficient, PMC-a activation decreases quickly (as abstract rule information becomes less important), paralleled by little additional increase in PMC-p engagement, reflecting that the pragmatic representation has quickly reached a stable state. Interestingly, based on the median split data, the PMC-p activation was significantly higher for high as compared with low performance improvement already in the initial practice phase at stimulus repetitions 1 and 2 (t = 2.36; P < 0.03). This finding suggests that high performance improvement during practice might be due to an already strong pragmatic code right from the outset of practice. In turn, such an interpretation implies that a stable pragmatic code might have been formed already during the instruction phase. In other words, high performance improvement during physical practice might be related to some form of mentally simulated practice already during the instruction phase, possibly similar to processes described in the context of motor imagery (Jeannerod and Frak 1999; Jeannerod 2001).
Relationship between Instruction-Related Activation and Behavioral Performance Improvement
The “simulated practice” interpretation proposed above is corroborated by findings from our final analysis step, which addressed the relationship between brain activation during the instruction phase and performance improvement during the subsequent practice phase. This analysis was performed to pursue the more general question how participants manage to implement newly instructed S–R mappings from the first trial, that is, without actual behavioral practice and solely based on reading and encoding the instructions. The general expectation was that a better encoding of S–R rules during the instruction phase, as indicated by a stronger engagement of the relevant brain regions, would lead to a stronger performance improvement during the subsequent practice phase. This predicted positive relationship was found in multiple brain areas (Table 2, Fig. 5). Most interestingly, this included the same right lateral PMC–PFC region that exhibited the complex relationship between practice-related dynamics of BOLD activation and performance improvement visualized in Figure 6. For this region, a positive correlation with performance improvement was found primarily for transient instruction-related activation (see results reported in Table 2) but at a more lenient threshold of P < 0.01 (uncorrected) also for sustained instruction-related activation. This finding nicely supports the interpretation developed above, which posits 1) that the formation of pragmatic task representations relies on a complex interplay between anterior and posterior subregions within the lateral PMC–PFC and 2) that this work can be accomplished during the instruction phase via some form of simulated practice.
|Brain region||MNI coordinates||Statistics|
|x||y||z||BOLD (instruction) × performance improvement||BOLD (practice) × performance improvement|
|Brain region||MNI coordinates||Statistics|
|x||y||z||BOLD (instruction) × performance improvement||BOLD (practice) × performance improvement|
Note: Asterisks denote significant t-values and Rho values: *P < 0.05, **P < 0.01, ***P < 0.001, uncorrected at the voxel level. Abbreviations: R, right; L, left; mMFG, midportion of the middle frontal gyrus.
Significant activations with P < 0.05, corrected at the cluster level (i.e., at least 26 contiguous voxels). Rho denotes Spearmans correlation coefficient. All reported correlations for BOLD (instruction) × performance improvement are based on the BOLD estimate for transient instruction-related activation.
Importantly though, there were other brain regions that exhibited a positive correlation between instruction-related activation and practice-related performance improvement but did not at the same time also show a correlation between practice-related activation and performance (for a representative brain region, see Fig. 5B). Such a pattern was observed in several brain regions selectively for transient instruction-related activation, including the midportion of the middle frontal gyrus bilaterally and the pre-SMA, though the latter region failed to reach significance based on the Spearman rank correlation. Note that additional caution seems advised as all 3 regions failed to reach the cluster-level-corrected significance criterion equivalent to 26 contiguously activated voxels. This distinct correlation pattern suggests that these regions might not be directly involved in the transfer of control from abstract to pragmatic rule representations as would have been suggested by an additional correlation between practice-related activation and performance. Rather, these regions might be related to the encoding of abstract S–R rules into working memory in a way that is suited to successfully guide future behavioral rule implementation. In other words, the more effort is put into abstract rule encoding during the instruction phase by these regions, the easier becomes the retrieval and implementation during subsequent practice. With regard to the pre-SMA, this interpretation seems in line with a recent study showing stronger pre-SMA activation for merely instructed but not yet implemented S–R rules compared with not-instructed stimuli (Brass et al. 2009).
Finally, we also observed 2 regions, the right caudate nucleus and the pSPL, which exhibited the opposite pattern, that is, correlations between practice-related activation dynamics and practice-related performance improvement (see above) in the absence of a significant correlation between instruction-related activation and performance improvement (Fig. 5C). Thus, different from the lateral PMC–PFC region, the caudate nucleus and pSPL might only be involved in pragmatic encoding processes when true behavioral feedback is generated, that is, during physical practice but not during simulated practice.
The present study aimed to identify the neurocognitive basis of instruction-based learning involving novel arbitrary S–R mappings. Two key features of our study design stick out in comparison to previous research, including 1) the examination of rapid adaptation processes across only a few practice trials and 2) the explicit instruction of S–R rules as compared with trial-and-error learning procedures. We reasoned that a sequence of 2 core processes might be central to the successful implementation of newly instructed S–R rules. First, we postulated that the instructions would initially (i.e., before their first behavioral implementation) be encoded on an abstract, possibly verbal level. Second, for successful behavioral implementation, such an abstract rule representation was hypothesized to be transformed into a pragmatic representation. We reasoned that this transformation process would necessarily start right from the first implementation of each newly instructed S–R association (otherwise no behavior would be possible), followed by further representational strengthening across practice. In other words, with each consecutive behavioral implementation of a given S–R association during the practice phase, the pragmatic representation should become increasingly independent from the scaffolding by the abstract rule representation, which, in turn, should become less and less relevant. Thus, as a result of this incremental transfer of control we expected 1) behavioral performance to improve, 2) brain areas involved in the pragmatic implementation of S–R rules to become increasingly engaged, and 3) brain areas involved in a more abstract representation of rules to become increasingly disengaged.
A first important aspect of the results is that all 3 indicators of this postulated rapid transfer of control were indeed observed, including 1) the improvement of behavioral performance across the practice phase, 2) an activation increase in the pre- and postcentral gyri and in the striatum, and 3) an activation decrease in the LPFC and along the IPS. This overall pattern of practice-related activation dynamics seems to be at least in part consistent with findings from other neuroimaging studies examining the learning of arbitrary S–R mappings by trial and error (Deiber et al. 1997; Toni et al. 2001; Eliassen et al. 2003; Boettiger and D'Esposito 2005; Brovelli et al. 2008). Thus, despite the apparent procedural differences between instruction-based learning and trial-and-error learning, on a general level, some neurocognitive commonalities seem to exist. An in-depth conceptual analysis regarding the possible functional commonalities and differences among these 2 forms of learning is certainly not warranted based on the current study results alone. Yet, the striking activation decrease in LPFC and parietal cortex observed in the present study is reminiscent of findings from the trial-and-error learning literature reporting strong PFC activation early during learning (Eliassen et al. 2003; Brovelli et al. 2008) which decreases with learning progress (Deiber et al. 1997; Law et al. 2005). These similar activation dynamics might reflect common underlying working memory processes related to the maintenance, monitoring, and selection among S–R rules (Bunge et al. 2003; Postle 2006). As elaborated in more detail at the end of the Discussion, the current results suggest that such executive control processes (especially those supported by the PFC) seem to be disengaged extremely quickly across only a few practice trials when S–R rules are explicitly instructed and sufficiently simple. Furthermore, our findings of practice-related activation increases following the instruction of novel abstract S–R rules are, on a general level, in line with previous reports in both humans and nonhumans, showing activation increases or arguing for the importance of sensorimotor areas (Deiber et al. 1997; Boettiger and D'Esposito 2005; Zach et al. 2008) and the striatum (Winocur and Eskes 1998; Hadj-Bouziane et al. 2003; Boettiger and D'Esposito 2005) during visuo-motor learning tasks.
Importantly though, despite these similar patterns of results, different from trial-and-error learning procedures, the present study suggests that these activation dynamics (both BOLD increases and decreases) are probably not so much reflecting the extraction of novel S–R rules (i.e., figuring out which response belongs to which stimulus) but might rather be related to the formation and strengthening of pragmatic representations of S–R rules. This conclusion extends insights from previous trial-and-error learning studies, which could not disentangle activation changes due to rule extraction processes from those related to the improvement of rule pragmatics. By contrast, in the present design, S–R rules can be encoded successfully already during the explicit instruction phase, and thus, practice-related activation dynamics are likely related to the improvement of rule pragmatics rather than to rule extraction. This conclusion seems to be supported also by the striking lack of reliable practice-related activation dynamics in mediotemporal lobe (MTL) structures in the present study. The MTL has typically been implicated in studies of rule learning, be it arbitrary S–R rules (Yanike et al. 2009), stimulus classification rules (Poldrack et al. 2001), or even implicit rule acquisition which does not rely on explicit feedback processing as in trial-and-error learning (Rose et al. 2002; Schendan et al. 2003). Yet, different from the present study, all these paradigms required the extraction of rules in the first place. Thus, together with the present results, this suggests a specific role of learning-related MTL activation for rule extraction processes.
Maybe the most important and certainly the more specific finding of the present study, however, is that practice-related performance improvement was significantly correlated with practice-related BOLD activation dynamics exclusively in 3 brain regions: the lateral PMC–PFC, the caudate head, and the pSPL (though this latter region failed to reach significance on the cluster level). Because actual overt behavior should be a direct reflection of pragmatic rule implementation, this correlation strongly suggests that practice-related activation dynamics in the identified brain regions might indeed reflect an involvement in the formation of pragmatic representations. Such a conclusion seems particularly warranted with regard to the lateral PMC–PFC, which is exactly the region that we predicted to be involved in the rapid formation and stabilization of pragmatic S–R associations. The involvement of the lateral PMC–PFC in pragmatic rule formation as implicated by the present study results further specifies and differentiates the functional role of this region in the implementation of novel S–R rules as discussed previously (Petrides 2005, 2008). Going beyond existing observations, our study design enabled us to demonstrate this region's extremely rapid adaptation dynamics across only few practice trials. Moreover, our results suggest a functional subdivision into a more anterior, as compared with a more posterior, PMC subregion, which seems consistent with complementary roles during the transfer from an abstract code (more anterior PMC) toward a self-sufficient pragmatic code (more posterior PMC). On a more general level, this functional segregation is consistent with previous theoretical considerations suggesting that more cognitive, or abstract action plans might be supported by anterior PMC subregions, as compared with more concrete motor programs supported by posterior PMC subregions (Picard and Strick 2001).
A further analysis of the lateral PMC–PFC region provided strong evidence that the transfer of control from abstract to pragmatic rule representations takes place already before the first behavioral implementation of novel S–R associations. This conclusion is suggested by the finding that stronger activation of the lateral PMC–PFC during the instruction phase was associated with stronger performance improvement during the subsequent practice phase. Thus, it seems that simulated practice performed during the instruction phase—at least by some subjects—can to some extent contribute to the formation of pragmatic S–R representations, which would otherwise be left for the “real” practice phase. Interestingly, the caudate nucleus does not seem to be involved in such type of simulated practice as it did not show a reliable correlation between behavior and instruction-related activation. Still, the caudate nucleus, like the lateral PMC–PFC, seemed to be involved in the formation of pragmatic S–R representations during “physical” practice as indicated by a significant correlation between behavior and practice-related activation.
The practice-related activation increase in the caudate nucleus following explicit rule instruction seems at odds with previous results from studies that employed probabilistic stimulus classification tasks in which participants had to learn to classify exemplar stimuli into different stimulus categories. Specifically, it has been shown that the caudate nucleus is involved in classification learning only when the rules had to be extracted based on trial-and-error learning but not during the explicit instruction of identical rules (Poldrack et al. 2001; Shohamy et al. 2004, 2006). Yet, other studies suggest that the MTL, but not the caudate nucleus, seems to be related to the extraction of complex rules (Rose et al. 2002, 2004). In these experiments, the caudate nucleus exhibited an activation increase associated with the gradual “proceduralization of simple S–R associations” based on explicitly instructed behavioral rules, which seems similar to the practice of instructed S–R rules in the present study, though on a much slower timescale. Together, across different study types, the exact functional role of the caudate nucleus during learning and practice seems to remain a matter of future research, which needs to further clarify the relevance of rule-type (e.g., S–S vs. S–R), specific learning procedure, and possibly also the timescale of the examined adaptation dynamics. The present study contributes to this quest by demonstrating that the caudate nucleus exhibits a rapidly increasing engagement during the practice of novel, explicitly instructed, and simple S–R rules.
Finally, our results might have interesting implications with regard to notions on the relationship between PFC function and automatization. A classical conceptual framework that might come to mind first, suggests that novel and unpracticed tasks rely on some form of “supervisory attentional system” (i.e., high demands on executive control) localized in the PFC before behavior becomes more automatic and less dependent on this brain structure (Norman and Shallice 1986). Importantly though, such an account of PFC function is typically thought to refer to the transition from controlled to automatic information processing across an extensive practice phase much longer than in the present study (Schneider and Chein 2003; Chein and Schneider 2005). Yet, our results seem to suggest that the engagement of large parts of the PFC decreases extremely quickly down to baseline level within only a few practice trials. This observation suggests that the implementation of novel behavioral rules depends on the PFC only during a very brief initial period and more generally, that the time window of mandatory PFC recruitment might be much smaller than implied by classical notions of controlled versus automatic processing, at least in case of relatively simple S–R rules.
We wish to thank Kerstin Raum for managing the recruitment of participants and for collecting the data. Conflict of Interest: None declared.