Assessing the Neural Basis of Uncertainty in Perceptual Category Learning through Varying Levels of Distortion

The ability to divide objects and events into separate categories is a basic skill of living organisms and provides the basis for the efficient processing of a large number of environmental stimuli. It involves the act of responding differently to stimuli in separate categories, which makes it inevitable to come to distinct decisions, often in the presence of uncertain information. Here we define uncertainty as a psychological state in which a person has insufficient knowledge about what outcome will follow from what choice, that is, when the category membership of a stimulus is ambiguous.

Although the exploration of category learning has a rich history in cognitive sciences (see, e.g., Ashby & Maddox, 2005), relatively little is known about the precise neural mechanisms mediating the processing of uncertainty during the formation of new perceptual categories. In a variety of tasks that involve decision uncertainty, brain areas predominantly in the medial wall of the frontal cortex, including Brodmann areas 6, 8, 24, and 32, were consistently activated. The putative function of these brain areas has been suggested in cognitive processes such as error and performance monitoring and signaling need for behavioral adjustment (Ridderinkhof, Ullsperger, Crone, & Nieuwenhuis, 2004; Ridderinkhof, van den Wildenberg, Segalowitz, & Carter, 2004). Furthermore, findings from neurophysiological and neuroimaging studies indicate that the lateral prefrontal areas, the anterior insular cortex, and the intraparietal area are not only associated with increased attentional and working memory demands during the processing of uncertainty but also directly involved in the computation of a decision variable (Heekeren, Marrett, & Ungerleider, 2008). When uncertainty develops with the accumulation of information toward a decision, activation associated with it was observed in pFC and posterior parietal cortex (PPC; Huettel, Allen, & McCarthy, 2005), whereas Grinband, Hirsch, and Ferrera (2006) report a network including the medial frontal gyrus, the anterior insula, the thalamus, and the ventral striatum, showing increasing activation with uncertainty while subjects judged the length of line segments.

Considering the growing interest in situations in which a task concurrently activates more than one response tendency, current research has also begun to investigate the role of uncertainty during classification learning. Yet most of these studies used probabilistic stimulus-outcome contingencies to elicit uncertainty on the response level. Aron et al. (2004) investigated neural correlates of uncertainty in a probabilistic classification task and reported mainly midbrain activity in relation to the degree of uncertainty, whereas in a study of Seger and Cincotta (2005), the caudate nucleus and extrastriate visual areas were more active for probabilistic than random and deterministic stimuli. However, activity in the left superior frontal gyrus encompassing portions of the medial wall areas (BA 8 and 9) showed a reversed relationship. The authors discussed this finding in terms of a potential association between this frontal activity and classification uncertainty.

Because of the lack of empirical data, it is an open question whether the presented neural processes, especially those mediated by posterior medial prefrontal and posterior parietal areas (Huettel et al., 2005; Ridderinkhof, Ullsperger, et al., 2004), also play a role in categorization tasks in which uncertainty emerges not because of probabilistic stimulus response associations but at other levels of processing, for example, at the level of visual processing (Botvinick, Jonathan, & Cameron, 2004). It is increasingly becoming apparent that processes involved in perception (i.e., the way objects are represented by vision) and those involved in cognition (i.e., the way objects are identified and categorized) are nonseparable processes (Palmeri & Gauthier, 2004). Therefore, the goal of the present study is to examine the processing of uncertainty during a distinct type of categorization learning that is known to have a strong perceptual component, prototype distortion learning. It refers to acquiring the ability to sort items consisting of distortions of prototypical stimuli into categories and implies becoming proficient in the interpretation of a wide range of sensory inputs in terms of a limited number of abstracted categories (Seger et al., 2000).

One of the classical paradigms to investigate prototype distortion learning, which also will be used within the present experiment, is the dot pattern classification task (Posner, Goldsmith, & Kenneth, 1967; Posner, 1964). However, because of differences in the actual implementation of this classical task in neuroimaging studies, for example, concerning the number of categories to be distinguished or the presentation of feedback, results on its neuronal underpinnings have to be interpreted with care. Although it has been postulated that learning to identify distortions of a single prototype is rather a form of purely perceptual learning than higher order learning (Ashby & Ell, 2001), results on prototype distortion learning involving two or more categories indicate that working memory and visuospatial attention play a crucial role (Casale & Ashby, 2008; Little, Shin, Sisco, & Thulborn, 2006; Little & Thulborn, 2006; Ashby & Maddox, 2005; Vogels, Sary, Dupont, & Orban, 2002; Seger et al., 2000). Also, as the importance of subcortical contributions to cognitive processes is beginning to be addressed, the involvement of BG structures, especially the caudate nucleus, in prototype distortion learning has to be specified. The caudate has been shown to play a relevant role in procedural learning, and its implication in other forms of category learning is well established (Ashby & Ennis, 2006), whereas fMRI experiments on prototype distortion learning using observational versions of the task found no BG activation (Reber, Gitelman, Parrish, & Mesulam, 2003; Aizenstein, MacDonald, & Stenger, 2000; Reber, Stark, & Squire, 1998a, 1998b). However, observational tasks were also shown not to involve the cortico-striatal loop in other domains of category learning (Poldrack et al., 2001), whereas results on caudate activation from more recent investigations on feedback-based versions of prototype distortion tasks are contradictory. Although Little et al. (2006) observed activation in the caudate tail during classification, Vogels et al. (2002) did report a locus in the neostriatum only at a liberal threshold, and Seger et al. (2000) did not report any striatal activations.

The main goal of the present study was to investigate the neural network that is involved in perceptional decision making in the presence of systematically varied categorization uncertainty. We analyzed, using fMRI, the neuronal activation in a prototype distortion task with the explicit instruction to learn the distinction between two different categories of distorted dot patterns in the presence of trial-by-trial feedback. This makes the dot pattern classification task more similar and therefore more comparable with tasks used in other domains of category learning than some of the previous designs. A parametric manipulation of uncertainty allowed us to both assess areas involved in the processing of uncertainty and to discriminate networks supporting category learning from those specific to more general task-related processes (e.g., non-categorization-related attention and motor planning) as proposed by Grinband et al. (2006).

A further goal was to investigate changes in neural activity with regard to learning-dependent decreases of categorization uncertainty. To be able to draw conclusions about the learning process, we developed two parallel versions of the task, one of which was trained before the experiment. This approach provided the possibility to present the trained before the untrained task version (Boettiger & D'Eposito, 2005) and thereby eliminate some of the confounding effects of time and familiarity inherent in protocols where the same task is compared on the same participants before and after training (Poldrack, 2000). Moreover, we deliberately separated stimulus presentation from feedback by adequate time intervals to be able to differentially study categorization processes, that is, uncertainty and learning-related processes as well as feedback-related processes.

Previous work on category learning indicates that extensive learning of visual patterns is associated with activation decreases in a network, including regions involved in visual processing, object recognition, and spatial attention without the recruitment of additional areas in skilled performance (Little & Thulborn, 2006; Little, Klein, Shobat, McClure, & Thulborn, 2004). We expected to disclose decreases in this network, too, but hypothesized that these decreases would be predominantly attributable to the decreasing level of conflict and diminishing need for cognitive control in the course of successful learning. It was therefore predicted that the network showing learning-dependent decreases in activation would largely overlap with the areas identified as decision related by the uncertainty modulation. The neural network underlying decision-related processing in the context of categorization was shown to include frontal medial wall areas, lateral prefrontal, parietal, and anterior insular cortex as well as areas involved in visual processing including primary visual cortices and inferior temporal areas (Grinband et al., 2006; Little & Thulborn, 2006; Little et al., 2006; Huettel et al., 2005). These areas were expected to show increased activation with regard to an increase in categorization uncertainty, whereas areas supporting more general demands of the task like areas supporting general attention and the motor response should not be affected.

Concerning feedback processing, recent studies indicate that the mesencephalic dopaminergic system with its projections to striatal, medial prefrontal, and orbito-frontal areas plays a central role in feedback-based category learning (Daniel & Pollmann, 2010; Little et al., 2006; Seger & Cincotta, 2005; Aron et al., 2004). As a mechanism for this contribution, phasic activation changes coding for errors in reward prediction (Schultz, 2002) have been suggested. Because the presence and the time course of feedback were shown to impact different kinds of category learning differentially (Maddox, Ashby, & Bohil, 2003; Ashby, Maddox, & Bohil, 2002), we analyzed the processing of feedback specifically for prototype distortion learning. Here, we expected positive feedback to be associated with relative activation increases in subcortical dopaminergic structures (Delgado, 2007). Consistent with its role in performance monitoring, we anticipated relatively increased activation in the medial pFC in association with negative feedback (Yeung, Holroyd, & Cohen, 2005; Ridderinkhof, van den Wildenberg, et al., 2004).

Twenty-eight subjects (age range = 20–28 years, M = 23.96 years, SE = 0.43 years, 15 women) recruited from the Friedrich Schiller University community participated in the experiment. One of them was excluded from further analysis because of excessive head movement. None of the subjects reported a history of drug abuse and chronic bodily or neurological and psychiatric diseases, and all were without pathological findings on the Symptom Check List (SCL-90-R; Franke, 1995). The participants were right-handed according to the modified version of the Annet handedness inventory (Briggs & Nebes, 1975) and reported normal or corrected-to-normal vision. Informed written consent was obtained in accordance with the protocols approved by the ethics committee of the University of Jena before the experiment, and all subjects received an allowance of €10 per hour in return for their participation.

Four prototypes were generated by pseudorandomly distributing nine filled white dots on a black background within a matrix of 30 × 30 cells. This matrix was the center portion of a 50 × 50 grid, which was sufficient to include all distortions of the prototypes, that is, all exemplars of the category (Posner, 1964). To avoid salient dot pair features the minimal distance between two dots needed to be at least three cells (Vogels et al., 2002) and to prevent subjects from adopting simple heuristics on the basis of the spatial distribution of the dots, we required the centroid for each prototype to lie within in the central 10 × 10 cells of the grid. The resulting four patterns represent the designated prototypes A, B, C, and D.

All subjects were trained on the task in two runs directly before scanning either with stimuli from categories A and B or from categories C and D. The selection of the category pair to be trained was counterbalanced across subjects, and training was performed in a separate quiet room. The training stimuli were different from the testing stimuli but derived in the same way from the same prototypes. Subjects were instructed that they would now see several dot patterns and that their task was to learn to classify the patterns according to their similarity into two groups on the basis of the feedback they would be given after each trial. To shorten the training time, the first training session was self-paced with seven training exemplars for each category and level of distortion, each presented twice, resulting in 84 randomized trials. The second training session showed training stimuli in a pace comparable with the fMRI testing with 42 training trials. After completing these two training sessions, subjects were placed in a comfortable position in the MRI scanner, and their head was positioned on a vacuum pillow between padded clamps to prevent movement.

Experimental stimuli were presented on a mirror positioned on the head coil, which reflected pictures projected onto a screen at the head end of the scanner by a computer video projector (model GT1150; NEC Corporation, Tokyo, Japan). Behavioral responses were collected through an MRI-compatible fiber-optic button box (LUMITouch; Photon Control Inc., Burnaby, BC, Canada), with two keys positioned next to the participant below the right hand. Participants were instructed to respond using their right index and middle fingers. A third training session identical to the second one was run before the start of the experiment. In the first session of the experiment, subjects had to categorize 42 dot pattern stimuli from the trained categories presented in pseudorandomized order. Each dot pattern stimulus was presented for 4 sec, and the participants were instructed to press the response button within this time frame as soon as they were reasonably confident about their decision. To increase sensitivity, we introduced a variable delay of 2 to 6 sec (M = 4 sec). Subsequently, feedback consisting of the words “Right!,” “Wrong!,” or “Missed!” and an indication of the correct category, for example, “Category A,” appeared on the screen for 1.5 sec. After an interval of 7 to 11 sec (M = 9 sec), the next dot pattern stimulus was presented. After the first session was completed, a second followed. It was identical to the first one but with stimuli of the untrained category pair. Each session had a duration of approximately 17 minutes.

Functional images were recorded on a 3-T whole-body scanner (MAGNETOM Trio, A Tim System; Siemens, Erlangen, Germany) with a 12-channel head matrix coil by using a whole-brain T2*-weighted EPI sequence, covering a volume of 40 parallel slices with 3.0-mm slice thickness (with a distance factor of 10%) and a voxel size of 2.7 × 2.7 × 3.0 mm (repetition time = 2040 msec, echo time = 26 msec, flip angle α = 90°, and field of view = 192 × 192 mm, matrix 72 × 72). Parallel imaging (GRAPPA) with an acceleration factor of two and 30 reference lines was used. Two runs, each with 440 volume acquisitions, were collected from each participant.

To exclude images contaminated by T1 saturation effects, we discarded the first four EPI images of each run from further analysis. The functional images were preprocessed with the statistical parametric mapping software (SPM5; Wellcome Department of Cognitive Neurology, London, UK). Preprocessing included slice timing correction using the middle image of the volume as reference slice and 3D motion correction, that is, rigid body realignment to the mean of all images. We ensured that head movement was below 3 mm and 3° for each participant. Images were normalized to MNI space (Evans et al., 1993) using the standard EPI template of SPM5. The data were spatially smoothed using a Gaussian filter of 8 mm FWHM, and a temporal high-pass filter of 128 sec was applied to remove low frequency confounds.

The first-level analysis of the imaging data was conducted using an event-related design. Regressors of interest represented the three levels of distortion of the categorization task, positive as well as negative feedback and the delay period between task and feedback. To control for the potential confound of subjects' movement, we entered individual movement parameters as covariates into the design matrix as estimated during the realignment step. All regressors were convolved with a model of the HRF. Contrast images for the three classification tasks against baseline were calculated separately for each session and subject. Because of too few error trials in most subjects in the trained session, the feedback phase contrasts of positive versus negative and negative versus positive feedback were only calculated in the untrained session. All contrasts were submitted to the second-level group analyses with subject as the random effect variable.

For the analysis of the classification task, a 3 × 2 ANOVA with the factors Level of Distortion and Training was performed. Two different contrasts were of interest: the training effect (untrained > trained) and the effect of uncertainty (parametric modulation of the effect of categorization by the level of distortion of the stimuli). For the analysis of the effect of feedback, the two individual t contrasts were submitted to a one-sample t test.

Except where indicated, activations were identified as significant if they passed a height threshold of p < .05 corrected for multiple comparisons at the whole-brain level using the false discovery rate correction (Genovese, Lazar, & Nichols, 2002) and a spatial extent threshold according to the number of expected voxels per cluster.

With increasing level of distortion, RTs were slowed, F(1.24, 31.04) = 47.71, p < .001, and error rates increased, F(1.66, 41.40) = 49.61, p < .001. A main effect of Training revealed that RT was shorter, F(1, 25) = 43.75, p < .001, and error rates were lower, F(1, 25) = 23.05, p < .001, in the trained (RT: M = 1573.90 msec, SD = 416.62 msec; error rates: M = 0.09, SD = 0.06) than that in the untrained run (RT: M = 2077.68 msec, SD = 566.79 msec; error rates: M = 0.22, SD = 0.13). No significant main effect for order or any interactions were observed for both measures. Error rate data are illustrated in Figure 2.

To observe the neuronal network implicated in decision making under uncertain conditions, we systematically varied the distortion of the stimuli and identified the areas showing an increasing response with increasing uncertainty while subjects categorized stimuli. As we had predicted, prefrontal areas, mainly the posterior medial prefrontal wall, the ventrolateral pFC (VLPFC), and the anterior insulae were activated. However, also strong activations along the dorsal and the ventral paths in both superior parietal and inferior temporal areas were observed as well as activations in the FEFs, right parahippocampus, thalamus, and both the ventral and the dorsal striata. Results are summarized in Table 2A and Figure 3, and bar charts illustrating the increase of percent of signal change with level of distortion are available in the Supplemental materials. Inclusive and exclusive masking on the basis of a threshold of p < .05 with the training effect showed that except for the right lateral orbito-frontal (Tal: x = 45, y = 49, z = −10) locus and the right parahippocampal gyrus all identified areas overlapped with or directly bordered on those identified to decrease with training.

To examine differences between the trained and the untrained performance of the task, we compared two parallel task versions, one of which was trained before scanning, within a single fMRI session. No areas showed higher activations in the trained as compared with the untrained run. Areas showing higher activations in the untrained as compared with the trained run are summarized in Table 2B and Figure 3. Inclusive and exclusive masking of this contrast with the effect of uncertainty showed that both contrasts activated similar networks. All clusters of activation overlapped with those reported for the effect of uncertainty; however, the focus of the premotor activations was observed to be more laterally located, and both dorsal and ventral striatal activations were more extended.

The effect of feedback valence was examined by comparing positive with negative feedback in the untrained run. As predicted, areas in the medial prefrontal wall responded more to negative as compared with positive feedback. As shown in Figure 4A, the locus of this activation was anterior to those observed to be activated during decision making. In addition, in this contrast, a fronto-parietal network was activated along with the inferior temporal areas, the precuneus, and a cluster in the posterior cingulate cortex. All activations are listed in Table 3A. At the chosen threshold, no areas showed higher activations in response to positive as compared with negative feedback. Because of our strong a priori hypothesis, we conducted this analysis at the less conservative threshold of p < .001 (uncorrected). As shown in Figure 4B at this threshold, activations were observed in the right ventral striatum and bilaterally in the hippocampus. All activations are listed in Table 3B.

The aim of the current study was to specify the network subserving prototype distortion learning under high levels of uncertainty. To this end, we conducted an fMRI prototype learning experiment that manipulated task difficulty by applying three different mathematically defined levels of distortion to the prototype. In addition, two parallel tasks were administered, one of which was trained before the experiment. Results indicate that the brain networks showing training-related decreases in activation, and those showing uncertainty-related increases largely overlap. They include the posterior parietal and inferior temporal areas, the FEFs, posterior medial prefrontal areas, the anterior insula, and a cortico-striatal–thalamic loop through the body of the caudate nucleus. It is important to note that varying the level of distortion of the stimuli does not specifically affect one level of processing but generally increases task demands. Therefore, the applied paradigm is suited to specify the whole neural network that is relevant to solving the task but not to distinguish between conflicts at, for example, the perceptual or motor level. Combining the present results with previous data, however, leads to testable hypotheses concerning the specific contributions of the different parts of the network supporting prototype distortion learning.

A large part the observed network, namely, posterior parietal areas along the intraparietal sulcus and inferior temporal areas as well as the FEFs, are commonly reported as the neural substrates of visuospatial processing (visual association cortices), spatial attention (parietal lobules), and the control of eye movements (FEFs) (Rosano et al., 2002; LaBar, Gitelman, Parrish, & Mesulam, 1999; Kanwisher, Chun, McDermott, & Ledden, 1996). Because of its complex multimodal responses, the PPC is often viewed as association cortex that integrates information from multiple senses (e.g., Culham & Kanwisher, 2001). Within the visual system, regions of the PPC are traditionally supposed to form a major component of the dorsal stream (Farivar, 2009), which is associated with the encoding of information for visually guided actions. In nonhuman primates, the lateral intraparietal area, which is supposed to be equivalent to the human intraparietal sulcus, is known to be involved in decision making and contains neurons encoding category membership on the basis of motion direction (Freedman & Assad, 2006). Given the fact that in the experiment by Freedman and Assad (2006) lateral intraparietal activation was not modulated by task difficulty, Grinband and Ferrera (2006) conclude that parietal areas are not involved in comparing the sensory data against the category boundary. This argument is supported by the fact that Grinband and Ferrera found no modulation of posterior parietal activity by categorization uncertainty in an fMRI experiment with humans. However, their experiment used simple one-dimensional lines that are not likely to activate areas involved in visuospatial processing. In the current experiment, complex two-dimensional stimuli were used, and category membership was defined entirely via spatial configurations. Under these conditions, a strong modulation of intraparietal activation by categorization uncertainty was observed, which indicates the participation of the PPC in categorization processes in this special case. The literature on decision-making under uncertainty provides some evidence on the function of the PPC in this context. Several studies reported rising PPC activation with increasing uncertainty on the stimulus-outcome level (Huettel et al., 2005; Volz, Schubotz, & von Cramon, 2003, 2004). Huettel et al. (2005) suggest that given the role of the PPC in the generation and modification of context-appropriate responses, its increased activity with uncertainty is the neuronal correlate of a more difficult and more attention-demanding generation process.

In addition, a cortico-striatal–thalamic loop through the body of the caudate observed to show increased activation with increasing task demands. As part of the visual cortico-striatal–thalamic loop with inferior temporal areas, the body of the caudate nucleus has been associated with implicit category learning (Seger, 2008). Although its role is well established for other forms of category learning (for an extensive review, see Ashby & Ennis, 2006), the caudate's association with prototype distortion learning is less clear. Prototype distortion learning has been suggested to rely on perceptual learning in tasks where only one category is presented, and subjects have to decide if stimuli are members of the category or not (A/non-A version), whereas behavioral data predict that in tasks where two categories have to be dissociated (A/B version), other memory systems are recruited (Casale & Ashby, 2008). At least two studies on prototype distortion learning with feedback, one using an A/B version (Seger et al., 2000) and the other one using a hybrid form (A/B/none) (Vogels et al., 2002), report no significant striatal activations. However, missing activation in the study of Seger et al. (2000) might be due to the use of a 1.5-T scanner, which is known to have a limited power at imaging deep brain structures (Poldrack & Willingham, 2006), whereas Vogels et al. (2002) did report a focus in the anterior neostriatum when using a less conservative threshold (p = .052). In contrast, Little et al. (2006) observed BG activations in a feedback-based version of the dot pattern paradigm, and Zeithamova, Maddox, and Schnyer (2009) observed stronger activations in the body of the caudate for a A/non-A classification task compared with an A/B classification task. Our results confirm the involvement of subcortical structures in A/B prototype distortion learning and can be interpreted within the framework proposed by Cincotta and Seger (2007). They suggest a central role to the body and tail of the caudate nucleus for the representation of associations between stimuli and categories. Increasing the level of distortion of the stimuli will render this comparison more demanding, whereas with training the comparison process requires less resources.

Activations that are directly related to task difficulty, that is, increase with uncertainty and decrease with training, were also observed in the insular cortices and the inferior and ventrolateral pFC. The inferior pFC or the VLPFC has repeatedly been found to increase in activation with increasing uncertainty (Schlösser et al., 2009; Koch et al., 2008). Likewise, several studies have reported anterior insular involvement in decision making under uncertainty (Grinband et al., 2006; Huettel et al., 2005; Paulus, Rogalsky, Simmons, Feinstein, & Stein, 2003; Volz et al., 2003; Critchley, Mathias, & Dolan, 2001). A framework for understanding the activations observed in the present experiment is given by Singer, Critchley, and Preuschoff (2009). They propose that the anterior insula is involved in learning about risk and uncertainty associated with a given decision and integrates this information with representations of bodily and affective states in dependence on the individual risk preference and appraisal of the context. An interpretation of our results based on this model is that the higher risk of negative reward outcomes and the emotional and somatic states invoked by it are reflected in insular activations.

A further important finding of the current study is an increasing posterior medial prefrontal activation with increasing task difficulty in a visual classification task. This area has been shown to be consistently involved in tasks requiring conflict monitoring, whereas it is still a matter of debate at which levels of processing it responds to conflict (Botvinick et al., 2004). At least one study directly compared conflicts at the stimulus evaluation level with response conflicts using a version of the Eriksen flanker task (van Veen, Cohen, Botvinick, Stenger, & Carter, 2001) and found that posterior medial prefrontal areas only monitor the presence of response conflicts, whereas others showed that activity can be induced in situations where no motor response was required (Monchi, Petrides, Petre, Worsley, & Dagher, 2001; Robertson et al., 2000). The major source of uncertainty in this study certainly lies within the perceptual domain at the level of stimulus evaluation. However, because an explicit motor response was required, the possibility cannot be ruled out that the observed activation arose not during stimulus evaluation but during the preparation of the motor response.

To investigate the neural substrates of feedback processing, we also compared the functional activation in response to positive compared with negative feedback. Activations that were stronger for negative feedback mainly included posterior medial and right lateral prefrontal areas. This is in line with a large body of research showing that medial prefrontal areas are activated by response errors and negative feedback. It was proposed that this information is then conveyed to lateral prefrontal areas, indicating the need to increase the level of adaptive cognitive control (for a review, see Ridderinkhof, van den Wildenberg, et al., 2004). Activations that were stronger for positive than negative feedback were only observed at a less conservative threshold (p < .001) and included the right ventral striatum as well as both hippocampi. The role of the ventral striatum as part of the dopaminergic system in the processing of positive feedback is well established (Delgado, 2007). However, the observed parallel hippocampal activation is in contrast to some previous results. Although two studies even reported a negative correlation between medial-temporal and striatal activations (Seger & Cincotta, 2006; Poldrack et al., 2001), others observed no hippocampal activations (Seger & Cincotta, 2002, 2005). On the other hand, at least two studies also reported simultaneous activations as they were observed in this experiment (Cincotta & Seger, 2007; Little et al., 2006). Seger and Cincotta (2005) used a paradigm in which subjects had to categorize balanced random stimuli and also found higher hippocampal activations for positive than for negative feedback. Yet they did not observe this difference in probabilistic and deterministic conditions. Therefore, a tentative explanation of the current results is that subjects recruit the explicit hippocampal memory system only for very difficult or even random tasks. For rule-based category learning, this interpretation was proposed theoretically by Ashby and Valentin (2005) and was later on empirically examined by Nomura et al. (2007). They observed higher hippocampal activations in correct trials (which included positive feedback) in a rule-based but not an information–integration task. The results of this experiment indicate that feedback processing in classification learning might both depend on striatal and hippocampal contributions.

The current investigation reports the whole breadth of the network subserving prototype distortion learning. Activations with an increasing level of distortion of the stimuli include visual association and posterior parietal cortices, the anterior insula, the posterior medial pFC, and the body of the caudate nucleus. These structures might reflect uncertainty at several stages of the decision process. On the basis of previous findings, the interpretation is suggested that the PPC activation reflects the more attention-demanding evaluation of the stimuli (Huettel et al., 2005). The body and the tail of the caudate as part of a visual cortico-striatal–thalamic loop have been proposed to be directly related to the representation of associations between stimuli and visual categories (Seger, 2006), indicating that the observed activation is attributable to a further level of processing, that is, the more complex comparison process of the current stimulus to the category. Further, the anterior insular activation is suggested to reflect negative states induced by the higher risk of receiving negative feedback (Singer et al., 2009). The posterior medial pFC has been previously observed to be activated by decision uncertainty and has been proposed to reflect conflicting motor responses. If this is the case, prior stimulus evaluation processes in the current experiment were not able to resolve the conflicting response tendencies until the motor response is planned. Because this experiment did manipulate uncertainty on all of the levels presented above, these interpretations remain tentative and serve to provide hypotheses that can be tested in future investigations. Present data also illustrate that in the context of prototype learning, a decrease in task demands allows a reduction of activation in task-relevant regions. Most importantly, a practice-related decrease in task demands reduces activation in similar networks as does the decrease in task demands caused by the systematic variation of task difficulty. We were able to show that in prototype distortion learning, a form of learning with a strong perceptual component, the dopaminergic system is activated during feedback processing. Also, recent experiments on reward-related influences on sensory decision making and perception (Law & Gold, 2009; Pleger et al., 2009) reveal that dopaminergic processes are able to modulate sensory processes. Taken together, these results indicate that reward-related processes play a central role in prototype distortion learning.

The decision to assign a stimulus to a certain category in perceptual category learning involves several subprocesses, including signal detection, comparison of the stimulus to an abstract category boundary, prototype or other seen exemplars, motor planning, and outcome evaluation (Grinband et al., 2006) in addition to perceptual learning. By parametrically varying the level of distortion and comparing trained with untrained performance, the current study was not only able to specify the network subserving prototype distortion learning in general but also to show that it is directly related to decision making and responds differentially to the level of decision uncertainty in the task. The reported neural network of prototype distortion learning includes medial prefrontal, parietal, and occipito-temporal areas as well as a cortico-striatal–thalamic loop through the body of the caudate nucleus.

Reprint requests should be sent to Reka Daniel, Department of Psychiatry and Psychotherapy, Friedrich Schiller University of Jena, Jahnstrasse 3, 07743 Jena, Germany, or Department of Experimental Psychology, Otto von Guericke University Magdeburg, Universitätsplatz 2, 39106 Magdeburg, Germany, or via e-mail: rdaniel@ovgu.de.