Abstract

We used event-related fMRI to identify the brain regions engaged during explicit and implicit sequence learning (ESL and ISL, respectively). Twenty-four subjects performed a concurrent ESL and ISL task. Behavior showed learning in both conditions. Prefrontal (PFC), striatal, anterior cingulate cortex (ACC) and visual regions (V1, V2 and V3) were engaged during both ESL and ISL. With ESL there was increased activity in the visual regions on the predictable (i.e. learned pattern) trials. With ISL, however, there was a relative decrease in activity in visual regions. The opposite patterns in the visual regions highlight the different effects of ESL and ISL. The learning process was distinguished from the result of learning, by fitting subjects’ functional magnetic resonance imaging data to their learning curve. This analysis revealed more extensive PFC activity during ESL and caudal ACC activity specific for the result of learning analysis, when the expected response was violated. Our results suggest a relative dissociation of the brain regions engaged during ESL and ISL, whereby ESL and ISL can be viewed as partially distinct but overlapping parallel processes.

Introduction

It has previously been demonstrated that individuals can learn about the sequential structure of events without being explicitly aware of this structure. In the classic serial reaction time task (SRTT; Nissen and Bullemer, 1987), for instance, individuals become faster at responding to simple reaction time events that follow a predetermined repeating sequence. This has been shown to occur even when the subjects are not explicitly aware that a sequential pattern exists (Willingham et al., 1989). In certain circumstances, however, explicit processing does seem to interfere with implicit learning. For instance, Nissen and Bullemer (1987) showed that instructing subjects to count auditory tones, concurrent with the sequence learning task, eliminated the implicit sequence learning effect. In subsequent studies, Schmidtke and Heuer (1997) argued that the interaction of implicit and explicit learning was specific for task integration. Their studies showed that explicit processing interfered with the implicit learning only insofar as there was a decrease in the predictability of the implicit rule due to integration with the other features. More recently, in a series of studies, Jimenez and Mendez (1999), has shown that implicit and explicit sequence learning can proceed concurrently without interference. This was found specifically to be the case with probabilistic sequence rules.

In the present functional magnetic resonance imaging (fMRI) study we use concurrent implicit and explicit probabilistic variants of the SRTT to test hypotheses about the brain regions involved during concurrent implicit and explicit sequence learning. We hypothesized that the implicit and explicit sequence learning circuits would be dissociable. We predicted that as in previous implicit sequence learning studies (Berns et al., 1997; Rauch et al., 1997) we would find striatal activation specific for the implicit condition and, as in previous explicit sequence learning studies (Sakai et al., 1998) prefrontal activation specific for the explicit learning condition. Additionally, we hypothesized that the implicit and explicit conditions would differ in terms of their activation in the visual cortex. Although, not commonly reported in functional imaging studies of the SRTT, other implicit versus explicit learning studies such as studies of priming and category learning have reported increased activation in the extrastriate cortex with explicit learning and relative decreases in activation with implicit learning. We predicted that if we used an analysis approach similar to those studies we would also find a difference between implicit and explicit sequence learning in the extrastriate cortex in the current study

Several recent studies (Rauch et al., 1995; Willingham et al., 2002; Schendan et al., 2003) have also contrasted implicit and explicit sequence learning using versions of the SRTT. These studies report mixed results. In a 15O-CO2 positron emission tomography study, Rauch et al. (1995), report striatal activation specific for the implicit condition and increased frontal activation in the explicit condition. In block design fMRI studies, Willingham et al. (2002) and Schendan et al. (2003) both report a significant overlap between the implicit and explicit conditions with striatal and frontal activation during both implicit and explicit block types. The current study differs from the previous in that the current design is event-related (D’Esposito et al., 1999) and thus is potentially more sensitive to the activity on individual learning trials, rather than being influenced by the state of learning induced by the block type.

Previous imaging studies of learning (Gabrieli et al., 1997; Buckner, 1998; Buckner and Koutstaal, 1998; Reber et al., 1998; Sakai et al., 1998; Schacter and Wagner, 1999; Aizenstein et al., 2000) have contrasted learned versus non-learned trials. This is also a primary contrast in the current study. However, recent theories of sequence learning have also emphasized the dynamic process of learning. For instance, early in the course of the sequence all of the trials are novel and the subject has little basis to predict the next trial. However, as learning progresses a structured learned rule develops, which the subject can use to predict the next trial, either explicitly or implicitly. This rule can then be followed or violated leading to possible modification of the learned rule.

Parallel distributed processing models of sequence learning, such as those described by Cleeremans and McClelland (1991) and also Berns and Sejnowski (1998), formalize the distinction between the process and the result of learning; there is a learning rule in the network that is distinct from the feed-forward decision making role in the network. The roles interact but are conceptually different and have different time-courses of activation. In the current study we draw on insight from these models and separately analyze the fMRI time-series to distinguish the process from the result of learning.

Materials and Methods

Event-related fMRI was done during a concurrent implicit and explicit sequence learning task. Twenty-six subjects participated. Fourteen subjects participated in the initial study design, described below, and referred to as study 1. The remaining subjects participated in a slightly revised version of the task (study 2). This revised design took advantage of an upgrade in the MR scanner software that allowed for faster image acquisition. Except where noted, the design of study 1 and study 2 were identical.

Research Participants

All participants were right-handed, 18–25 years old, and were recruited by advertisement in a community newsletter. One participant was excluded because they were uncomfortable in the scanner and chose not to complete the protocol, and two subjects were excluded because of equipment failure. Of the 14 participants in study 1, seven were men and seven were women, with a mean (± SD) age of 23.9 ± 5.1 years. Each individual was paid $50. Of the 10 participants in study 2, five were men and five women, with a mean age of 22.3 ± 1.7 years.

Procedure

This experiment was divided into two different conditions: pre-practice and sequence learning. The sequence learning task involved concurrent implicit and explicit components. The task is a variation of the SRTT. The standard SRTT involves several hundred trials that follow a repeating sequence and then a block of randomly ordered trials. In the current study the goal was to optimize the contrast between the learning and random trials, therefore we used a modification of the SRTT in which the pattern and random trials were interspersed, similar to that described by Howard and Howard (1997). Moreover, since the study used a rapid event-related design it was important to randomize the occurrence of pattern and random trials, therefore we used probabilistic sequences (such as described by Jimenez and Mendez, 1999).

Stimuli

For each subject eight different sequences of 85 colored shapes were generated. Three colors (red, green and blue) and three shapes (circle, square and triangle) were used. The sequence of colors and shapes were determined independently using two different first-order markov chains (i.e. the probability distribution of states at time i + 1 is completely determined by the state at time I; see Fig. 1). At each state in the markov chain the next color had a 70% chance of following in the set order (e.g. red, green, blue) and a 30% chance of violating the set order (i.e. either repeating or skipping a position). Similarly, the next shape had a 70% chance of following a set order of shapes and a 30% chance of violating the order. The markov chains were thus used to generate a random sequence of colored shapes for each subject, such that the color and shape followed independent probabilistic sequences; correlations between the color and shapes sequences were low (maximum r = 0.18).

Pre-practice

Prior to performing the category learning tasks, each subject was pre-practised on the mapping of color to button press. This pre-practice occurred just before entering the MRI machine. The subjects spent 7 min responding to patches of red, green and blue by pressing the left button, under their index finger, for red, the middle button, under their third finger, for green, and the right button, under their ring finger, for blue.

Sequence Learning

Each stimulus appeared in the center of the screen. The trials occurred every 3.5 s in study 1 and every 2 s in study 2. After 750 ms the red, green or blue colored stimulus changed to white, i.e. a white circle, square or triangle on a black background. The subject was instructed to look at each colored shape and respond to the color as quickly and accurately as possible by pressing the button corresponding to the color. The white shape stayed on the screen for an additional 2000 ms (750 ms for study 2). A fixation stimulus was then displayed on the screen for 750 ms (500 ms for study 2), until the onset of the next trial.

Participants were instructed that while they were to respond as quickly as possible to the color, they were also to look for a sequential pattern in the shapes. They were told that the sequential pattern did not always hold, but that they should look for one and report this after each of the learning blocks. As described, the shapes did follow a sequence determined by the Markov-chain (Fig. 1a). Learning the sequence in shapes was the explicit learning component of the task. Unbeknownst to the subjects, however, the colors also followed a sequential pattern, independent of the pattern of shapes. Learning the sequence of colors was the implicit sequence learning component. Subjects’ change in reaction-time and accuracy at identifying the color was used to measure implicit learning. We assessed for explicit knowledge of the color sequence at the completion of the task. Subjects were asked a series of increasingly specific questions, such as, did they notice anything about the colors and did they notice a sequence in the colors. Finally they were told that there was a sequence in the colors and were asked if they were aware of the sequence. We had the color disappear after 750 ms to encourage subjects to respond quickly to the color, pilot testing showed that when the color remained for the entire stimulus duration some subjects would respond more slowly to the color, apparently as they thought about the shape.

Stimulus presentation and response recording was controlled by a Power Macintosh computer, using the PSYSCOPE software package (Cohen et al., 1993). The stimuli subtended ∼30° of the visual field. Responses were measured using a fiber optic button box attachment. This equipment provided one msec temporal resolution. It was strapped to the participant’s right hand. Participants were instructed to press these buttons in response to the stimuli.

Scanning Procedures

MRI scanning was conducted on a 1.5 T General Electric Signa Scanner. Structural MRI was performed prior to the functional scans. These were used to align the functional MRI acquisition, and also for cross-registration of the functional scans for the planned group analysis. The structural scans were acquired as T1 weighted images. Thirty-six 3.8 mm slices were obtained in the oblique axial plane, aligned within the anterior cortex–posterior cortex (AC–PC) line.

The functional images had in plane resolution of 64 × 64, with 3.75 × 3.75 mm pixels. Images were acquired in eight blocks of 85 trials each over the course of learning. The presentation of stimuli was synchronized with the scanning such that one image was acquired per trial. Since the trials occurred more frequently than the 12–14 s hemodynamic response, this is rapid event-related fMRI design (e.g. Dale and Buckner, 1997). Rapid event-related studies generally use fixation intervals to allow the hemodynamic response to return to a baseline, thus improving the estimate of the hemodynamic response (HDR) for a given trial condition. Because fixation trials would disrupt a sequence of stimuli, it is unclear how the addition of fixations would affect the cognitive process of sequence learning. Thus, in the current study we chose to avoid the fixations. Instead, to mitigate against the possibility of saturation of the BOLD signal, we relied on the ‘random’ intermixing of trial types. Previous studies (e.g. Dale and Buckner, 1997) have shown that the HDR summates approximately linearly at an interval of 2 s. Thus convolving an idealized HDR with a trial type vector (as was done in the current study) should correlate with the fMRI signal, and allow for the general linear model analysis described in our report.

Study 1 Scanning Parameters

Functional scanning was performed using a two-shot spiral pulse sequence, with TE = 35 and TR = 1750. Twenty-two contiguous 3.8 mm slices were obtained in the oblique-axial plane. The second most inferior slice was aligned on the AC–PC line.

Study 2 Scanning Parameters

Functional scanning was performed using a one-shot spiral pulse sequence, with TE = 35 and TR = 2000. Twenty-six contiguous 3.8 mm slices were obtained in the oblique-axial plane. The fourth most inferior slice was aligned on the AC–PC line.

Image Analysis

For each subject, movement correction was performed using a six-parameter linear algorithm (six-parameter AIR; Woods et al., 1998), the fMRI data was then spatially smoothed (using an 8 mm full width–half maximum Gaussian filter), normalized and detrended. The structural scans for all 24 subjects were registered to the first structural scan in this study. This was done using a 12-parameter linear algorithm (12-parameter AIR) that aligned the structural scans to a reference brain (the first subject of our study). The analysis for significant differences in activations was then performed voxel-wise, using the general linear model implemented in the AFNI fMRI analysis package (Cox, 1996).

Partial correlation maps for individual subjects were generated to reflect the extent to which voxels’ activity conformed to an a priori hemodynamic response function [double gamma function (Boynton et al., 1996; Dale and Buckner, 1997)]. For each subject four different partial correlation maps were generated: one map each for implicit and explicit learning, for both the result of learning (learned) and for the learning process. For the implicit learned analysis, the variance associated with the implicit learning and the explicit conditions was covaried out. This was done using multiple regression with three regressors: one for the implicit learned condition, one for the implicit learning condition and one for the explicit condition — this combination of regressors has relatively low collinearity (average partial R2 = 0.1 in multiple regressions of the regressors). The residual variance associated with just the implicit learned regressor was used to generate an individual subject’s r-map. Each of the three other r-maps was generated similarly.

To generate the partial correlation maps we first constructed the separate regressors for the learned and learning conditions in both the implicit and explicit case. The regressors were constructed based on both stimulus order and performance [the difference in reaction time (RT) between rule-following and rule-violating trials for the implicit rule, and questioning about the explicit rule]. The performance was used to estimate a linear learning curve (the ramp function), which was then array-multiplied by a 680 item binary vector of the stimulus order (1 for each implicit pattern trial, 0 elsewhere) and convolved with the a priori hemodynamic response. For instance, the result of implicit learning regressor was generated using a ramp function with a positive slope, which modeled the increased activation over time that is predicted to occur as the result of learning. The implicit learning-process regressor was generated using the same initial 680-item binary vector as the result of implicit learning regressor, except that for the learning process this vector was multiplied by a ramp function with a negative slope, to model the decrease in activity that is predicted to occur as learning proceeds and there is a diminishing difference between the actual pattern and the learned pattern. For the explicit condition, similar learned and learning-process regressors were generated.

Implicit and explicit regressors were also generated without using the ramp function for use in the multiple regressions of the alternative condition. The implicit and explicit learning ramp functions were estimated based on the block number at which subjects demonstrated some learning of the rule. In the explicit case, this was the first block number after which the subject reported the correct sequence. In the implicit case this was the first block number in which the median reaction-time for non-pattern was greater than the median reaction-time for pattern trials. The ramp functions were then constructed as simple linear functions, capped at a maximum of 1 and a minimum of 0, and intersecting the learning-process-rate at the first block that showed learning (see Fig. 2). Sequence learning in the serial reaction time task seems to have a nearly linear learning rate, according to several studies with a similar design as the current task (Howard and Howard, 1997; Jimenez and Mendez, 2001). Thus we used the simple linear ramp function as described.

The resulting partial correlation maps were then used to do a single group t-test testing the null hypothesis that the mean r = 0. The statistical maps were thresholded using a cluster-size algorithm (Forman et al., 1995). A per-voxel alpha value of 0.005, t(23) > 3.1, was used, and as in similar fMRI analyses (Aizenstein et al., 2000; Barch et al., 2000), we chose a cluster-size threshold of eight voxels. We corrected for multiple comparisons using Monte-Carlo simulations, as implemented by AlphaSim (Cox, 1996) with an image-wise false-positive rate of 0.05. The regions were then overlayed onto the reference anatomical image, which was transformed to the Talairach atlas (Talairach and Tournoux, 1988). For each region of interest (ROI) identified in the partial correlation maps, we examined the average time-series in the pattern (P) versus the non-pattern (NP) trials and report whether P > NP (or NP > P) at the expected peak of the hemodynamic response, 4–6 s after trial onset. The multiple regression analysis finds for each condition the unique variance accounted for by that condition after the variance associated with every other condition has been removed. To more directly compare the implicit versus explicit, and learned versus learning conditions we also performed t-tests directly comparing the rs for each of the identified ROIs.

Results

Behavioral Performance

After each of the eight learning blocks we tested for explicit learning by asking subjects to report the observed sequence of shapes. All 24 subjects reported the correct explicit rule at the end of the first block.

The learning curve for the implicit condition, as reflected by reaction time and accuracy, is shown in Figure 3. Note that the learning curve shows a significant separation of pattern and non-pattern trials, but there appears to be a speed-accuracy trade-off. That is, during the blocks with low RTs for non-pattern trials (blocks 6 and 8), there is a corresponding decrease in accuracy. To show the overall learning curve, corrected for the speed-accuracy trade-off, we use Townsend and Ashby’s (1983) method of dividing RT by accuracy. This is shown in panel c of the figure. Additionally, we also show a bar chart of the overall accuracy and RT change. On the non-pattern trials, as compared to the pattern trials, there was a significantly greater reaction-time [matched sample t(23) = 5.21, P < 0.0001] and lower accuracy [matched sample t(21) = 2.74, P < 0.05, two-tailed]. On the non-pattern trials the mean reaction-time was 656.9 ± 114.4 and mean accuracy was 0.91 ± 0.06. On pattern trials, mean RT = 628.0 ± 126.6 and mean accuracy = 0.94 ± 0.05. (Accuracy data from two subjects was lost due to a computer malfunction.)

The non-pattern trials were composed of repeating and non-repeating colors. We contrasted the RT and accuracy data between these two trial types and found no significant difference in either case; for RT, matched sample t(23) = 0.97, and for accuracy, matched sample t(21) = 1.29. On the repeating colors, the mean reaction-time was 654.4 ± 105.8 and mean accuracy was 0.92 ± 0.05. On non-repeating trials, mean RT = 661.5 ± 126.60 and mean accuracy = 0.91 ± 0.08.

To test whether there was any explicit awareness of the implicit pattern subjects were debriefed as to whether they noticed any sequential pattern in the colors of the stimuli. All subjects responded ‘no’. Frequently the subjects spontaneously commented that they were too busy looking for a pattern in the shapes to think about a pattern in the colors. Thus the concurrent explicit condition may have helped to mask the implicit rule.

fMRI Results: Result of Learning Analysis

t-Tests of subjects’ r-maps were performed using subject as a random variable. Planned analyses were performed to test for a significant correlation of neural activity and the presence of pattern trials during both the implicit learning condition and the explicit condition. The neuroanatomic regions found to be significant in these analyses are identified in Table 1. Below we discuss in more detail the activations in the occipital cortex and the subcortical region (see Fig. 4).

Result of Implicit Learning

As predicted, the pattern of activity associated with the result of implicit sequence learning included the visual cortex and the striatum. The sign of the β in the visual particles was negative, suggesting a lower activation for the pattern as compared to the random trials. This result is consistent with previous studies of implicit learning in that activation for the implicitly learned stimuli shows a diminished response compared to the novel, non-learned stimuli. This extends previous work by showing this diminished response in a sequence learning task. Sequence learning is fundamentally different from previous implicit learning tasks that found diminished activation (e.g. priming and category learning); with sequence learning, the stimuli themselves are no different in the learning and random trials, rather it is the position in the sequence, or the context, that differs.

Result of Explicit Learning

The pattern of activity found with explicit sequence learning was similar to that in the implicit condition. The visual cortex, the caudate and the putamen showed activation in regions similar to those identified in the implicit case. The primary difference between the implicit and explicit conditions was the direction of activation change in the visual cortex. Whereas in the implicit condition the visual particle had a negative β, indicating a diminished response associated with the learning trials, the opposite pattern was found in the explicit case. The difference between implicit and explicit learning has been demonstrated previously in studies of category learning. As we predicted, this same difference was observed with sequence learning. Thus, showing that even when the stimuli themselves are similar in the learning and non-learning trials, implicit and explicit learning have opposite activation effects in the extrastriate cortex.

We had predicted that striatal activation would not be found in the explicit condition, but would be limited to the implicit condition. Previous functional imaging studies of sequence learning had found the striatal activation primarily in the implicit tasks, thus we hypothesized it was specific for the implicit condition. Contrary to our prediction, however, we found striatal activation in the explicit, as well as the implicit condition. This likely reflects the inherent overlap of implicit and explicit sequence learning, both in terms of the cognitive phenomena as well as the neural processes.

Process of Learning

The neuroanatomic regions found to be significant in these analyses are identified in Table 2. The activation patterns for both the implicit and explicit conditions are similar to the patterns found in the result of learning contrast, with one primary exception: frontopolar activation was specific for the process of learning analysis in the explicit condition. The increased frontal activation in the explicit condition supports the view that the effortful process of explicit learning engages additional prefrontal processes beyond those utilized during performance of the learned task.

Implicit Versus Explicit and Result Versus Process of Learning

t-Tests were done to directly compare the r’s in the implicit versus explicit conditions, and in the result versus the process of learning, for each of the identified ROIs in Tables 1 and 2. For most of the reported areas a significant difference was found between implicit and explicit learning (see Table 3). The only regions that did not show significant differences were the insula and the primary somatosensory cortex in the result of learning analysis. In the direct contrast of the result versus the process of learning, all the identified ROIs were found to be significant in this analysis, at P < 0.01.

Discussion

This study contrasted regional brain activity during implicit and explicit sequence learning. Our results support a multiple (Goschke, 1998), as opposed to a single system view (Perruchet and Amorim, 1992) of sequence learning. The different patterns of activation during implicit and explicit learning were not mutually exclusive, but rather overlapped, with common areas of increased activation in the basal ganglia. This common signal suggests that elements of these multiple systems can also operate in parallel. Our observation of an overlap in the fMRI signal associated with implicit and explicit sequence learning replicates recent findings by Willingham et al. (2002) and Schendan et al. (2003), which have also found overlap in the fMRI signal during implicit and explicit sequence learning. We extend the previous studies using an event-related design, and finding differences between implicit learning in the visual cortex, as well as differences between the result and the process of learning.

The result of learning and the process of learning were separately analyzed using the general linear model. In the result of learning analysis we found (i) implicit and explicit learning showed opposite patterns of activation in the visual cortex; (ii) there was generally more prefrontal activation in the explicit condition; and (iii) contrary to our hypothesis there was striatal activation in the explicit condition (in addition to the implicit condition). The effects in the visual cortex extend the previously observed differences in implicit versus explicit learning. Previous studies (Buckner and Koutstaal, 1998; Reber et al., 1998; Aizenstein et al., 2000) have demonstrated decreased activation in the visual cortex associated with priming and implicit category learning, and increases in activation associated with explicit category learning and recognition tasks. These findings have been interpreted as evidence of the different processes underlying implicit and explicit processing in the visual cortex. That is, implicit learning is associated with increased processing efficiency after repeated exposures. This increased efficiency leads to both a decreased fMRI signal, as well as a faster reaction time. Explicit processing is thought to engage posterior attentional mechanisms leading to increased activation. The current study extends these previous finding to sequence learning, which is fundamentally different from the other tasks in that the only difference between learning random stimuli is presentation order (not stimuli characteristics). Thus, our observation of a separation in the visual cortex between implicit and explicit sequence learning shows that learning context is sufficient (even in the absence of stimuli differences) to drive visual cortex changes.

The additional prefrontal activation with explicit rather than implicit learning supports previous studies that have emphasized the importance of prefrontal activation in explicitly controlled processes, such as working memory (Cohen et al., 1997; Smith and Jonides, 1998), selective attention (Carter et al., 1998; MacDonald et al., 2000) and problem solving (Koechlin et al., 1999). This also extends previous distinctions between implicit and explicit category learning (Reber et al., 1998; Aizenstein et al., 2000). The striatal activation during explicit learning is contrary to our initial hypothesis. We hypothesized that the striatum played a role specific for implicit learning. The fact that we find striatal activation in both conditions, suggests that either there is overlap in performance of the task (i.e. even though the instructions are explicit, there is still an implicit component of the task) or the neural processes of explicit and implicit learning both involve the striatum. Our study is not able to distinguish between these possibilities.

In the result of learning analysis we also found anterior cingulate cortex (ACC) activity during both the implicit and explicit learning conditions. Previous work (e.g. Carter et al., 1998) has suggested that the anterior cingulate may play a role in monitoring conflict during a response to a stimulus (i.e. increased activity during an incongruent trial of a Stroop task). The observation here that there is increased anterior cingulate activity on random (rather than pattern) trials of the learned sequence is consistent with the response conflict hypothesis for the ACC. That is, in the later trials, after the subject has learned the pattern, the random trials most violate the subject’s expected response and thus are associated with the greatest response conflict. Moreover, since this effect is seen in the implicit condition, it appears that the anterior cingulate responds to conflict even in the absence of awareness, as suggested by Berns et al. (1997).

This is the first functional imaging study that we are aware of that has modeled the process of sequence learning. Our results show that this approach is promising. In this case we identified more frontopolar activation involved in the explicit learning process analysis. This supports the view that the effortful process of explicit learning engages additional prefrontal processes beyond those utilized during decision making with a learned rule or during implicit learning.

Frontopolar activation has been found in a number of other functional imaging studies of learning, sequence processing, and hypothesis generation (Christoff and Gabrieli, 2000). The particular cognitive role of the frontopolar region may involve maintaining internal information, as suggested by Christoff and Gabrieli (2000). Alternatively, Koechlin et al. (1999) have described the frontopolar region as managing cognitive branching, i.e. actively maintaining information about a primary task while performing a subgoal. Our results are consistent with both the internal state and cognitive branching views. During the explicit sequence learning task the subjects maintain and manipulate their hypothesis of the shape sequence. Since the hypothesis is internally generated, this fits with Christoff and Gabrieli’s hypothesis. The explicit sequence learning task can also be thought as involving cognitive branching. A common strategy for explicit sequence learning involves re-evaluating on each trial the fit of different possible sequential pattern rules. Testing each of these rules can be thought as a evaluating subgoals while maintaining the overall goal of finding the right pattern, and thus is consistent with Koechlin et al.’s view.

Recently, several studies have examined the interaction of implicit and explicit learning during probabilistic category learning, using the weather task (e.g. Poldrack et al., 2001). These studies have identified competing neural systems during this task: the striatum shows increased activation and the medial temporal lobe shows decreased activation during implicit aspects of the task, whereas the opposite pattern is found during the explicit aspect of the task. Our results are similar to these in also showing striatal changes in the implicit condition. However, contrary to their findings we found striatal activation during the explicit task. Moreover, in our task we did not find significant medial temporal lobe changes.

A number of differences between the current task and the weather task might account for some the different findings. In the weather task individuals are asked to predict rain or sun based on a card, which predicts the correct answer between 60 and 80% of the time. In the implicit version subjects receive trial by trial feedback, whereas the explicit version is structured as a paired-associates learning task. This is fundamentally different from sequence learning, where the sequential relationship between the stimuli is learned rather than the individual features of the stimuli. It may be that sequential learning is inherently more ‘procedural’ and thus even the explicit version activates the striatum rather than the medial temporal lobe.

Alternatively, certain limitations of the current study might explain some of our findings. We made efforts to match the implicit and explicit versions of the learning tasks, but there are still significant differences in the relevant features and the response characteristics. Although unlikely, it could be that the differences between colors and shapes, rather than implicit versus explicit learning are responsible for the different patterns of activation found in this task. The different response characteristics for the implicit and explicit tasks were chosen to match the types of responses often used in previous studies of implicit or explicit tasks, i.e. reaction time for implicit and verbal report for explicit. Nevertheless, these differences do add another aspect, besides awareness, that distinguishes the tasks, and thus further limits our interpretation. Additionally, the early learning of the explicit rule, by the end of the first block in every subject, may have decreased our power to detect fMRI signal associated with the process of explicit learning, thus leading to a type II error in the explicit learning analysis. The primary results of the current study, however, do not suggest that this is a problem, i.e. we found more extensive activation in the explicit process of learning analysis.

Another limitation in the current study design is that the imaging parameters do not fully cover the medial temporal lobe. The bottom-most slice covers 3.8 mm below the AC–PC line, which excludes a large portion of the medial temporal lobe (MTL). In the current study we did not find significant MTL activation in the portion that was imaged. However, in previous studies MTL activation with explicit learning has been observed in regions centered below our imaging field. For example, in the weather task, the MTL activation was centered at 18 mm below the AC–PC line. Thus, we cannot draw conclusions from the lack of medial temporal lobe activation in the current study. A limitation of our fMRI analysis is the use of an idealized hemodynamic response function (as in SPM99, Wellcome Department of Cognitive Neurology). Using a single idealized HDR may decrease power as the HDR differs across subjects, as well as across brain regions within a subject (Buckner et al., 1996).

This study supports and extends current cognitive and neural theories of implicit and explicit sequence learning. The finding of different activation patterns in the visual cortex supports the multiple system theory of implicit and explicit learning. However, our findings that the systems are largely overlapping and proceed concurrently allow us to further refine hypotheses of implicit and explicit sequence learning. The overlapping pattern suggests that the dissociation is relative. That is, even though there are multiple learning systems these systems share much in common. Additionally, the concurrent nature of the two tasks suggests that they can complement one another. Implicit and explicit sequence learning thus take place in parallel and result in the development of complementary representations of sequential information. In the case of implicit learning this is evidenced by decreased cortical activation on the learned pattern and for explicit pattern it is evidenced by increased activation.

This research was supported by AAGP Pfizer/Eisai Fellowship (HJA), Hartford AFAR Fellowship (HJA), Burroughs Wellcome Translational Scientist Award (CSC), and NIMH grants T32-MH19986, R25-MH060473, and K02MH064190. We thank Angus MacDonald, Greg Siegle, Jonathan Cohen, Kate Fissell, and Victor Ortega for their assistance.

Figure 1. Markov chains that generate the color (a) and shape (b) sequences.

Figure 1. Markov chains that generate the color (a) and shape (b) sequences.

Figure 2. Examples of ramp functions used to generate regressors for fMRI analysis.

Figure 2. Examples of ramp functions used to generate regressors for fMRI analysis.

Figure 3. Learning curves during the implicit condition as reflected by reaction-time (a), accuracy (b), reaction-time/accuracy (c) and overall effect on reaction-time and accuracy (d).

Figure 3. Learning curves during the implicit condition as reflected by reaction-time (a), accuracy (b), reaction-time/accuracy (c) and overall effect on reaction-time and accuracy (d).

Figure 4. fMRI Results for result of learning analysis in explicit condition (a) and implicit condition (b).

Figure 4. fMRI Results for result of learning analysis in explicit condition (a) and implicit condition (b).

Table 1


 Talairach coordinates for all significant particles identified in the result of learning analysis

Region Coordinates P Voxels (nDirection 
Implicit learned     
Cuneus and precuneus (17, 18, 19) (–38, –69, 40) (7, –85, 10) <0.001 991 NP > P 
Striatum: cd, putamen (–17, 22, 10) (31, –35, 6)   0.036 358 NP > P 
Left posterior cingulate (30)  (–17, –66, 14)    0.009 190 NP > P 
Frontal     
 ACC (24, 32) (11, 26, 23) (–5, 26, 36) <0.001 356 NP > P 
 DLPFC (9.46) (43, 26, 32) (–54, 26, 27) <0.001 346 P > NP 
 Primary motor (4, 6) (–58, –11, 40) (–47, –7, 41) <0.001 528 NP > P 
 Primary sensory (1, 2, 3) (–58, –11, 44) (39, –28, 53) <0.001 190 NP > P 
 FEFs (BA 8) (–50, 11, 44) (47, 11, 44) <0.001 245 P > NP 
Insula (39, 17, 11) (–50, –17, 18) <0.001 125 NP > P 
Superior and inferior parietal (40, 7) (–58, –31, 48) (11, –73, 40)   0.009 278 NP > P 
     
Explicit learned     
Cuneus and precuneus (17, 18, 19) (–13, –85, 23) (19, –81, 28) <0.001 640 P > NP 
Striatum: cd, putamen (–9, 18, 2) (11, 22, 6)   0.003 433 P > NP 
Frontal     
 ACC (24, 32) (–1, 41, 10) (3, 37, 10) <0.001 420 NP > P 
 DLPFC (9. 44, 45, 46) (54, 26, 25) (52, 14, 19) <0.001 482 P > NP 
 Primary motor (4, 6) (–50, 3, 48) (39, –21, 53) <0.001 571 P > NP 
 Primary sensory (1, 2, 3) (–58, –11, 44) (39, –24, 53) <0.001 188 NP > P 
 Left FEF (BA 8) (–54, 7, 40) <0.001 275 P > NP 
Insula (–29, 22, 2) (43, –14, 15)   0.019 127 P > NP 
Inferior parietal (40) (–50, –31, 52) (39, –31, 53)   0.017  97 NP > P 
Region Coordinates P Voxels (nDirection 
Implicit learned     
Cuneus and precuneus (17, 18, 19) (–38, –69, 40) (7, –85, 10) <0.001 991 NP > P 
Striatum: cd, putamen (–17, 22, 10) (31, –35, 6)   0.036 358 NP > P 
Left posterior cingulate (30)  (–17, –66, 14)    0.009 190 NP > P 
Frontal     
 ACC (24, 32) (11, 26, 23) (–5, 26, 36) <0.001 356 NP > P 
 DLPFC (9.46) (43, 26, 32) (–54, 26, 27) <0.001 346 P > NP 
 Primary motor (4, 6) (–58, –11, 40) (–47, –7, 41) <0.001 528 NP > P 
 Primary sensory (1, 2, 3) (–58, –11, 44) (39, –28, 53) <0.001 190 NP > P 
 FEFs (BA 8) (–50, 11, 44) (47, 11, 44) <0.001 245 P > NP 
Insula (39, 17, 11) (–50, –17, 18) <0.001 125 NP > P 
Superior and inferior parietal (40, 7) (–58, –31, 48) (11, –73, 40)   0.009 278 NP > P 
     
Explicit learned     
Cuneus and precuneus (17, 18, 19) (–13, –85, 23) (19, –81, 28) <0.001 640 P > NP 
Striatum: cd, putamen (–9, 18, 2) (11, 22, 6)   0.003 433 P > NP 
Frontal     
 ACC (24, 32) (–1, 41, 10) (3, 37, 10) <0.001 420 NP > P 
 DLPFC (9. 44, 45, 46) (54, 26, 25) (52, 14, 19) <0.001 482 P > NP 
 Primary motor (4, 6) (–50, 3, 48) (39, –21, 53) <0.001 571 P > NP 
 Primary sensory (1, 2, 3) (–58, –11, 44) (39, –24, 53) <0.001 188 NP > P 
 Left FEF (BA 8) (–54, 7, 40) <0.001 275 P > NP 
Insula (–29, 22, 2) (43, –14, 15)   0.019 127 P > NP 
Inferior parietal (40) (–50, –31, 52) (39, –31, 53)   0.017  97 NP > P 

The table lists all particles that reached image-wise significance at threshold of P < 0.05 using Monte-Carlo simulations as implemented by AlphaSim (Cox, 1996). Talairach coordinates (Talairach and Tournoux, 1988) are given for the most significant voxel in each particle. All particles were bilateral except where noted (i.e. left posterior cingulated in implicit condition and left FEF in explicit condition).

Table 2


 Talairach coordinates for all significant particles identified in the process of learning analysis

Region Coordinates P Voxels (nDirection 
Implicit learned     
Cuneus and precuneus (18, 19) (–17, –88, 14) (15, –81, 32) <0.001  381 NP > P 
Frontal     
 DLPFC (9, 46) (43, 22, 36) (–54, 30, 18)   0.003  206 P > NP 
 Primary motor (4, 6) (–46,14,48) (39, 3, 49) <0.001  270 NP > P 
 Primary sensory (1, 2, 3) (–58, –14, 44) (39, –21, 53) <0.001   98 NP > P 
 FEFs (BA 8) (23, 30, 49) (–46, 22, 44) <0.001  121 NP > P 
Parietal (40, 7) (–26, –81, 31) (39, –31, 53)   0.002  176 NP > P 
     
Explicit learned     
Cuneus and precuneus (17, 18, 19) (–21, –88, 14) (15, 81, 32) <0.001 1415 P > NP 
Left striatum: cd, putamen (–17, 11, 2)    0.035  494 NP > P 
Posterior cingulate (30) (–5, –50, 6) (3, 50, 6) <0.001  205 NP > P 
Frontal     
 ACC (24, 32) (–17, 37, –6) (3, 24, 32) <0.001  334 NP > P 
 DLPFC (9, 44, 45, 46) (–54, 30, 18) (43, 22, 36) <0.001  544 P > NP 
 Primary motor (4, 6) (–46, 14, 48) (47, 7, 41) <0.001  741 P > NP 
 Primary sensory (1, 2, 3, 43) (–58, –14, 44) (15, –33, 53) <0.001  284 NP > P 
 FEFs (BA 8) (27, 26, 49) (–42, –18, 48) <0.001  290 P > NP 
 Frontopolar (10) (39,45,19) (–42, –45, 23) <0.001  163 NP > P 
Insula (–50, 11, 6) (48, –3, 7) <0.001  250 P > NP 
Temporal (22) (–58, 3, 5) (52, 1, 7) <0.001  431 NP > P 
Superior and inferior parietal (5, 7, 40) (–42, –39, 57) (19, 77, 32) <0.001  342 NP > P 
Thalamus (–1, –11, 6) (1, –11, 6)   0.003  370 NP > P 
Region Coordinates P Voxels (nDirection 
Implicit learned     
Cuneus and precuneus (18, 19) (–17, –88, 14) (15, –81, 32) <0.001  381 NP > P 
Frontal     
 DLPFC (9, 46) (43, 22, 36) (–54, 30, 18)   0.003  206 P > NP 
 Primary motor (4, 6) (–46,14,48) (39, 3, 49) <0.001  270 NP > P 
 Primary sensory (1, 2, 3) (–58, –14, 44) (39, –21, 53) <0.001   98 NP > P 
 FEFs (BA 8) (23, 30, 49) (–46, 22, 44) <0.001  121 NP > P 
Parietal (40, 7) (–26, –81, 31) (39, –31, 53)   0.002  176 NP > P 
     
Explicit learned     
Cuneus and precuneus (17, 18, 19) (–21, –88, 14) (15, 81, 32) <0.001 1415 P > NP 
Left striatum: cd, putamen (–17, 11, 2)    0.035  494 NP > P 
Posterior cingulate (30) (–5, –50, 6) (3, 50, 6) <0.001  205 NP > P 
Frontal     
 ACC (24, 32) (–17, 37, –6) (3, 24, 32) <0.001  334 NP > P 
 DLPFC (9, 44, 45, 46) (–54, 30, 18) (43, 22, 36) <0.001  544 P > NP 
 Primary motor (4, 6) (–46, 14, 48) (47, 7, 41) <0.001  741 P > NP 
 Primary sensory (1, 2, 3, 43) (–58, –14, 44) (15, –33, 53) <0.001  284 NP > P 
 FEFs (BA 8) (27, 26, 49) (–42, –18, 48) <0.001  290 P > NP 
 Frontopolar (10) (39,45,19) (–42, –45, 23) <0.001  163 NP > P 
Insula (–50, 11, 6) (48, –3, 7) <0.001  250 P > NP 
Temporal (22) (–58, 3, 5) (52, 1, 7) <0.001  431 NP > P 
Superior and inferior parietal (5, 7, 40) (–42, –39, 57) (19, 77, 32) <0.001  342 NP > P 
Thalamus (–1, –11, 6) (1, –11, 6)   0.003  370 NP > P 

The table lists all particles that reached image-wise significance at threshold of P < 0.05 using Monte-Carlo simulations as implemented by AlphaSim (Cox, 1996). Talairach coordinates (Talairach and Tournoux, 1988) are given for the most significant voxel in each particle. All particles were bilateral except where noted (i.e. left striatum in explicit condition).

Table 3


 Direct comparison of implicit and explicit correlations

Region Significance (P
Learned  
Cuneus and precuneus (17, 18, 19) 0.0073 
Striatum: cd, putamen 0.0069 
Posterior cingulate (30)   
Frontal  
 ACC (24, 32) 0.022 
 DLPFC (9. 44, 45, 46) 0.030 
 Primary motor (4, 6) 0.009 
 Primary sensory (1, 2, 3) 0.264 
 FEFs (BA 8) 0.038 
Insula 0.234 
Superior and inferior parietal (40, 7) 0.033 
  
Learning  
Cuneus and precuneus (17, 18, 19) 0.0097 
Left striatum: cd, putamen 0.025 
Posterior cingulate (30)  
Frontal  
 ACC (24, 32) 0.02 
 DLPFC (9. 44, 45, 46) 0.0015 
 Primary motor (4, 6) 0.0041 
 Primary sensory (1, 2, 3, 43) 0.0117 
 FEFs (BA 8) 0.0039 
Insula 0.022 
Superior inferior parietal (5, 7, 40) 0.012 
Temporal (22) 0.0174 
Thalamus 0.017 
Region Significance (P
Learned  
Cuneus and precuneus (17, 18, 19) 0.0073 
Striatum: cd, putamen 0.0069 
Posterior cingulate (30)   
Frontal  
 ACC (24, 32) 0.022 
 DLPFC (9. 44, 45, 46) 0.030 
 Primary motor (4, 6) 0.009 
 Primary sensory (1, 2, 3) 0.264 
 FEFs (BA 8) 0.038 
Insula 0.234 
Superior and inferior parietal (40, 7) 0.033 
  
Learning  
Cuneus and precuneus (17, 18, 19) 0.0097 
Left striatum: cd, putamen 0.025 
Posterior cingulate (30)  
Frontal  
 ACC (24, 32) 0.02 
 DLPFC (9. 44, 45, 46) 0.0015 
 Primary motor (4, 6) 0.0041 
 Primary sensory (1, 2, 3, 43) 0.0117 
 FEFs (BA 8) 0.0039 
Insula 0.022 
Superior inferior parietal (5, 7, 40) 0.012 
Temporal (22) 0.0174 
Thalamus 0.017 

The table shows the significance of a two-tailed paired t-Test (with 23 degrees of freedom) comparing the r’s in the regions identified in Tables 1 and 2.

References

Aizenstein H, Macdonald AW, Stenger VA, Nebes RD, Larson JK, Ursu S, Carter CS (
2000
) Complementary category learning systems identified using event-related functional MRI.,
J Cogn Neurosci
 
12
:
977
–987.
Barch DM, Braver TS, Sabb FW, Noll DC (
2000
) Anterior cingulate and the monitoring of response conflict: evidence from an fMRI study of overt verb generation.
J Cogn Neurosci
 
12
:
298
–309.
Berns GS, Cohen JD, Mintun MA (
1997
) Brain regions responsive to novelty in the absence of awareness.
Science
 
276
:
1272
–1275.
Berns GS, Sejnowski TJ (
1998
) A computational model of how the basal ganglia produce sequences.
J Cogn Neurosci
 
10
:
108
–121.
Boynton G, Engel S, Glover G, Heeger, D (
1996
) Linear systems analysis of functional magnetic resonance imaging in human V1.
J Neurosci
 
16
:
4207
–4221.
Buckner RL (
1998
) Functional-anatomic correlates of object priming in humans revealed by rapid presentation event-related fMRI.
Neuron
 
20
:
285
–296.
Buckner RL, Koutstaal W (
1998
) Functional neuroimaging studies of encoding, priming, and explicit memory retrieval.
Proc Natl Acad Sci USA
 
95
:
891
–898.
Buckner RL, Bandettini PA, O’Craven KM, Savoy RL, Petersen SE, Raichle ME, Rosen BR (
1996
) Detection of cortical activation during averaged single trials of a cognitive task using functional magnetic resonance imaging.
Proc Natl Acad Sci USA
 
93
:
14878
–14883.
Carter CS, Braver TS, Barch DM, Botvinick MM, Noll D, Cohen JD (
1998
) Anterior cingulate cortex, error detection, and the online monitoring of performance.
Science
 
280
:
747
–749.
Christoff K, Gabrieli JDE (
2000
) The frontopolar cortex and human cognition: evidence for a rostrocaudal hierarchical organization within the human prefrontal cortex.
Psychobiology
 
28
:
168
–186.
Cleeremans A, McClelland JL (
1991
) Learning the structure of event sequences.
J Exp Psychol Gen
 
120
:
235
–253.
Cohen JD, MacWhinney B, Flatt M, Provost J (
1993
) PsyScope: an interactive graphic system for designing and controlling experiments in the psychology laboratory using Macintosh computers.
Behav Res Methods Instrum Comput
 
25
:
257
–271.
Cohen JD, Perlstein WM, Braver TS, Nystrom LE, Noll DC, Jonides J, Smith EE (
1997
) Temporal dynamics of brain activation during a working memory task.
Nature
 
386
:
604
–608.
Cox RW (
1996
) AFNI: software for analysis and visualization of functional magnetic resonance neuroimages.
Comput Biomed Res
 
29
:
162
–173.
Dale AM, Buckner RL (
1997
) Selective averaging of rapidly presented individual trials using fMRI.
Hum Brain Mapp
 
5
:
329
–340.
D’Esposito M, Zarahn E, Aguirre GK (
1999
) Event-related functional MRI: implications for cognitive psychology.
Psychol Bull
 
125
:
155
–164.
Forman S, Cohen J, Fitzgerald M, Eddy W, Mintun M, Noll D (
1995
) Improved assessment of significant activation in functional magnetic resonance imaging (fMRI): use of a cluster-size threshold.
Magn Reson Med
 
33
:
636
–647.
Gabrieli JD, Brewer JB, Desmond JE, Glover GH (
1997
) Separate neural bases of two fundamental memory processes in the human medial temporal lobe.
Science
 
276
:
264
–266.
Goschke T (
1998
) Implicit learning of perceptual and motor sequences: evidence for independent learning systems. In: Handbook of implicit learning (Stadler MA, Frensch PA, eds), pp.
401
–444. Thousand Oaks, CA: Sage.
Howard JH, Howard DV (
1997
) Age differences in implicit learning of higher-order dependencies in serial patterns.
Psychol Aging
 
12
:
634
–656.
Jimenez L, Mendez C (
1999
) Which attention is needed for implicit sequence learning?
J Exp Psychol Learn Mem Cogn
 
25
:
236
–259.
Jimenez L, Mendez C (
2001
) Implicit sequence learning with competing explicit cues.
Q J Exp Psychol A Hum Exp Psychol
 
54A
:
000
–000.
Koechlin E, Basso G, Pietrini P, Panzer S, Grafman J (
1999
) The role of the anterior prefrontal cortex in human cognition.
Nature
 
399
:
148
–151.
MacDonald AW 3rd, Cohen JD, Stenger VA, Carter CS (
2000
) Dissociating the role of the dorsolateral prefrontal and anterior cingulate cortex in cognitive control.
Science
 
288
:
1835
–1838.
Nissen MJ, Bullemer P (
1987
) Attentional requirements of learning: evidence from performance measures.
Cogn Psychol
 
19
:
1
–32.
Perruchet P, Amorim M-A (
1992
) Conscious knowledge and changes in performance in sequence learning: evidence against dissociation.
J Exp Psychol Learn Mem Cogn
 
18
:
785
–800.
Poldrack RA, Clark J, Pare-Blagoev EJ, Shohamy D, Creso Moyano J, Myers C, Gluck MAL (
2001
) Interactive memory systems in the human brain.
Nature
 
414
:
546
–550.
Rauch SL, Savage CR, Brown HD, Curran T, Alpert NM, Kendrick A, Fischman AJ, Kosslyn SM (
1995
) A PET investigation of implicit and explicit sequence learning.
Hum Brain Mapp
 
3
:
271
–286.
Rauch SL, Whalen PJ, Savage CR, Curran T, Kendrick A, Brown HD, Bush G, Breiter HC, Rosen BR (
1997
) Striatal recruitment during an implicit sequence learning task as measured by functional magnetic resonance imaging.
Hum Brain Mapp
 
5
:
124
–132.
Reber PJ, Stark CE, Squire LR (
1998
) Cortical areas supporting category learning identified using functional MRI.
Proc Natl Acad Sci USA
 
95
:
747
–750.
Sakai K, Hikosaka O, Miyauchi S, Takino R, Sasaki Y, Putz B (
1998
) Transition of brain activation from frontal to parietal areas in visuomotor sequence learning.
J Neurosci
 
15
:
1827
–1840.
Schacter D, Wagner A (
1999
) Medial temporal lobe activations in fMRI and PET studies of episodic encoding and retrieval,
Hippocampus
 
9
:
7
–24.
Schendan HE, Searl MM, Melrose RJ, Stern CE (
2003
) An fMRI study of the role of the medial temporal lobe in implicit and explicit sequence learning.
Neuron
 
37
:
1013
–1025.
Schmidtke V, Heuer H (
1997
) Task integration as a factor in secondary-task effects on sequence learning.
Psychol Res
 
60
:
53
–71.
Smith EE, Jonides J (
1998
) Neuroimaging analyses of human working memory.
Proc Natl Acad Sci USA
 
95
:
12061
–12068.
Talairach J, Tournoux P (
1988
) Co-planar stereotaxic atlas of the human brain. New York: Thieme Verlag.
Townsend JT, Ashby FG (
1983
) The stochastic modeling of elementary psychological processes. Cambridge: Cambridge University Press.
Willingham DB, Nissen MJ, Bullemer P (
1989
) On the development of procedural knowledge.
J Exp Psychol Learn Mem Cogn
 
15
:
1047
–1060.
Willingham DB, Salidis J, Gabrieli JD (
2002
) Direct comparison of neural systems mediating conscious and unconscious skill learning.
J Neurophysiol
 
88
:
1451
–1460.
Woods RP, Grafton ST, Holmes CJ, Cherry SR, Mazziotta JC (
1998
) Automated image registration. I. General methods and intrasubject, intramodality validation.
J Comput Assist Tomogr
 
22
:
139
–152.