Dissociable Retrosplenial and Hippocampal Contributions to Successful Formation of Survey Representations

Introduction

The daily task of navigating a complex environment requires an ensemble of cognitive processes that rely on distinct spatial representations (Franz and Mallot, 2000). These representations can be assembled from direct experience during navigation or from indirect information such as verbal descriptions or pictorial maps. Whenever humans experience an environment from a ground-level perspective, the ensuing spatial representations may consist of both route knowledge (Gillner and Mallot, 1998; Mallot and Gillner, 2000; Wolbers et al., 2004) and survey knowledge (Thorndyke and Hayes-Roth, 1982; Ruddle et al., 1997), depending at least in part on the learning strategy applied (Kyllonen, 1993; Allen, 1999). Given that metric information about distance and direction can be incorporated early in the learning process (Montello, 1998), it seems possible to develop survey knowledge from the very beginning of a learning experience, contrary to hierarchical models of spatial learning (Siegel and White, 1975).

Although it remains unclear whether navigation in well known environments is based on unified cognitive maps (Wang and Brockmole, 2003; Wiener and Mallot, 2003), various neuroimaging and lesion data, along with the discovery of place cells in epileptic patients, led to the belief that allocentric representations preside in the human hippocampus (Ekstrom et al., 2003; Hartley et al., 2003; Iaria et al., 2003). Moreover, the retrosplenial cortex is thought to support stimulus conversion from egocentric reference frames in parietal cortex to allocentric reference frames in medial temporal regions and vice versa (Burgess et al., 2001a,b; Ino et al., 2002).

Because most studies on human navigation have focused on the retrieval of previously learned environments, our goal was to identify the neural structures involved in acquiring survey knowledge from ground-level navigation. To form such an observer-independent representation, spatial information initially encoded in egocentric coordinates has to be transformed into an allocentric reference system. This process requires the integration of self-motion cues, presumably involving retrosplenial and medial temporal cortices. Unfortunately, recent neuroimaging studies have not provided unambiguous findings regarding the precise functions of these areas. For example, the parallel between higher error rates/longer response times and increasing hippocampal activation during maze learning (Iaria et al., 2003) contradicts the observed positive correlation between navigational accuracy and hippocampal activation during retrieval of survey knowledge (Hartley et al., 2003). Although the experimental design of the former study did not permit dissociation of learning and retrieval, the results indicate that the hippocampus may be most important during the initial learning stages. As spatial representations become more and more accurate and performance approaches ceiling level, hippocampal contribution may decline, as has been shown for simple object-location tasks (Grön et al., 2001).

Here, we used our recently established virtual-reality paradigm (Wolbers et al., 2004), which provides behavioral parameters precisely indicating the emergence of survey knowledge. We used functional magnetic resonance imaging (fMRI) while participants repeatedly encountered a complex environment from a groundlevel perspective. Based on previous findings, we expected performance-dependent activation changes in retrosplenial and medial temporal cortices to correlate with behavioral data indicative of survey learning.

Materials and Methods

Participants. Initially, 17 healthy male volunteers with normal or corrected-to-normal vision gave written informed consent to participate in this study, which was approved by the local ethics committee. All subjects understood the instructions without difficulty, and none were aware of the hypotheses at the time of testing. Because six participants did not meet our behavioral criteria for survey learning (for details, see Results), the final data set comprised 11 subjects (age range, 19-28 years).

Experimental stimuli. We used Blitz3D (Blitz Research, Auckland, New Zealand) to animate a desktop virtual environment (Fig. 1), the layout of which was derived from the “hexatown” environment developed by Gillner and Mallot (1998). Brick stone walls on both sides of the road restricted the view to the immediate straight ahead; remote road sections or intersections were invisible. Twelve distinct buildings serving as landmarks were placed at four intersections (three landmarks per intersection). These buildings were hidden behind walls unless subjects were standing directly in front of them (see below). To obtain a control condition that was carefully matched for perceptual input, we designed a single corridor with varying landmarks placed behind walls at both ends of the corridor. The corridor comprised one road section and two intersections at both ends; its appearance, which was kept constant across sessions, was identical to the environment used for encoding.

Stimuli for the retrieval task consisted of snapshots of all buildings from the same viewpoints as those encountered during navigation. These images were of the original size; in addition, smaller versions of each image were also created. To test subjects' ability to retrieve information about spatial relationships between pairs of buildings, images were placed above one another, as shown in Figure 1 (bottom right). These pairs always consisted of buildings that were located at adjacent intersections in the environment; none of the trials depicted landmarks from the same or from remote intersections.

Procedure. Six learning and three control conditions were presented in separate sessions with a predetermined order. The experiment started with a control session, followed by three learning sessions, the second control session, an additional three learning sessions, and the final control session. Within each session, both the encoding and the retrieval tasks were presented with intermediary fixation periods (20 s). During encoding, participants were passively moved throughout the entire environment following the same route in all sessions. Subjects were instructed to infer the spatial layout of the environment and the correct landmark locations to ensure the best possible performance during retrieval; moreover, they were told that they would have to draw a map of the virtual world after fMRI scanning. The latter seems to be a very efficient method to induce survey learning according to repeated observations during preliminary behavioral testing.

As noted above, landmarks were hidden behind walls most of the time. However, when the midpoint of an intersection was reached, after a randomized delay between 500 and 800 ms, the wall in front disappeared, thereby unveiling the building behind. After 2500 ms, the subject (i.e., camera) performed a 60° turn (at 90.23°/s), and the wall was reinstated. After reaching a dead end, a 180° turn (at 90.23°/s) was executed, and the journey continued to the previously visited intersection. During control sessions, subjects were moved repeatedly back and forth along a single corridor. Close to each dead end, the wall would disappear after a randomized delay lasting between 500 and 800 ms to reveal one of the 12 buildings encountered during encoding. After 2500 ms, the camera performed a 180° turn (at 90.23°/s), the wall was reinstated, and the virtual movement was continued toward the opposite wall. Each time a wall vanished, a different landmark was seen, allowing for the presentation of all 12 of the buildings within six visits of each wall. The presentation order was randomized across subjects. Participants were told to pay attention to all buildings; no task was to be performed. Because we ensured that visual stimulation was carefully matched between control and encoding conditions, activation differences could reliably be attributed to the learning of the spatial layout.

During retrieval, pairs of original and small-sized landmark images were presented until a response was given. Participants were instructed to imagine that they were facing the large building and to assess the relative spatial position of the small building (referred to as the target building). This was done by pressing one of three response buttons (left button, target building located to the left; right button, target building located to the right; middle button, target building located behind). Note that the exact position of the target building within the adjacent intersection was irrelevant; the task only required the subject to determine the relative spatial position of the intersection containing the target building. Subjects were encouraged to respond as quickly as possible.

Eighteen pairs of buildings and six null events were presented in randomized order; intertrial intervals randomly varied between 3000 and 5000 ms. Six pairs depicted buildings encountered in immediate succession during encoding; these pairs are referred to as direct pairs (Fig. 1, pair A-B). Six other pairs contained landmarks that were not encountered in immediate order; however, the target building was located on an adjacent intersection that was visited during encoding immediately after the large building. These pairs are referred to as close pairs (Fig. 1, pairs A-C and A-D). The remaining six pairs differed from the close pairs in that the target building was taken from an adjacent intersection that was not visited immediately after the large building. These pairs are referred to as remote pairs (Fig. 1, pairs A-E, A-F, and A-G). To maximize the variance of the number of intermediate path segments, we included two close and two remote pairs in which the target building had been encountered before the large building. However, we ensured that none of these pairs presented the buildings from a direct pair in reverse order. In summary, although spatial distance was held constant across all pairs (the target building was always located on an adjacent intersection), the important difference lay in the varying temporal delays between encountering both landmarks during encoding. Whereas the buildings constituting direct pairs were separated by only one path segment during navigation, close and remote pairs contained buildings that were separated by varying numbers of up to 11 path segments.

The rationale behind this approach was to obtain a behavioral measure allowing for a clear identification of the nature of the acquired representation. The acquisition of route knowledge in the form of temporospatial associations between consecutive landmark views manifests itself in a significant behavioral advantage of direct pairs over both close and remote pairs (Wolbers et al., 2004), because only direct pairs contain landmarks encountered in immediate temporal succession during encoding. However, if subjects are able to infer a survey representation of the environment, no significant behavioral differences (response time and accuracy) among direct, close, and remote pairs would be expected, because spatial distances are identical.

The retrieval task during control sessions did not require the retrieval of spatial representations. In this condition, all 18 pairs used for retrieval, plus six additional pairs depicting the same building in regular and small sizes, were presented with a duration of 4000 ms along with six null events. Subjects were instructed to indicate by a button press whether both buildings were identical (right button) or different (left button). Intertrial intervals were identical to those in the learning sessions.

After fMRI scanning concluded, subjects were asked to draw a map of the environment as accurately as possible. To minimize possible interindividual variance related to the presence or absence of silent naming of the landmarks, participants had to memorize the names of all buildings before fMRI scanning. After presenting all 12 of the buildings accompanied by their names, subjects were asked to retrieve the name of each building; in cases of failure, the experimenter provided the correct name. This procedure continued until each participant could reliably name each landmark.

MRI acquisition. MR scanning was performed on a 3 T MRI scanner (Trio; Siemens AG, Munich, Germany) with a standard head coil. Thirty-seven contiguous axial slices (slice thickness, 3 mm) were acquired using a gradient-echo echo-planar T2^*-sensitive sequence (repetition time, 2.12 s; echo time, 25 ms; flip angle, 70°; matrix, 64 × 64; field of view, 192 × 192 mm).

A liquid crystal display video projector back-projected the stimuli on a screen positioned on top of the head coil. Subjects lay on their backs within the bore of the magnet and viewed the stimuli comfortably via a 45° mirror that reflected the images displayed on the screen. To minimize head movement, all subjects were stabilized with tightly packed foam padding surrounding the head. Encoding stimuli were presented using Blitz3D; retrieval stimuli were presented using Presentation (Neurobehavioral Systems, Albany, CA).

Image processing and statistical analysis. Image processing and statistical analysis were performed using SPM2 (www.fil.ion.ucl.ac.uk/spm/software/spm2). All volumes were realigned to the first volume, spatially normalized (Friston et al., 1995) to an echo-planar imaging template in a standard coordinate system (Evans et al., 1993), and finally, smoothed using a 9 mm full-width at half-maximum isotropic Gaussian kernel. To test the hypothesis of deducing survey knowledge from ground-level navigation, we modeled the onset of each encoding block as a δ function convolved with a hemodynamic response function (HRF). Specific effects were tested with appropriate linear contrasts of the parameter estimates for the HRF regressor. Data were analyzed for each subject individually (first-level analysis) and for the group. At the single-subject level, we applied a high-pass filter to remove baseline drifts. Design matrices containing nine separate sessions (six encoding and three encoding control sessions) were specified, thereby removing session-specific effects. In addition, we included realignment parameters to regress out movement-related activation. Separate contrast images for each of the nine regressors were subsequently entered into a random-effects analysis (Friston et al., 1999) to obtain results that could be generalized beyond the subjects taking part in this study.

A main effect of learning and the performance-related increase were assessed with a repeated-measures ANOVA; the problem of nonindependent data within subjects, as well as error variance heterogeneity, was addressed by performing a nonsphericity correction. To model a performance-related activation increase across encoding sessions, we first estimated the underlying learning curve for remote pairs by subjecting the mean performance data to the recently introduced state-space smoothing algorithm (Smith et al., 2004). This procedure was deliberately chosen because absolute performance values underlie random fluctuations that may not accurately represent the current state of learning. We specifically chose remote pairs because it is possible that, in addition to learning the layout of the environment, subjects acquire some knowledge about the route as well, which could influence performance, especially for direct pairs. In contrast, remote pairs contain buildings with very long temporal delays and are thus most difficult to solve by route knowledge. As a consequence, these pairs provide the purest measure of survey expertise that is available in our paradigm.

The normalized learning estimates, as obtained from the state-space smoothing algorithm, were then used as contrast weights. We also tested for a performance-related decrease in activation across encoding sessions, using the inverted normalized learning curve. Given the substantial variability across subjects regarding learning speed, we checked for activation predicting the learning-related change (i.e., the increment in performance from one session to the next) by applying the individual estimated learning curves for remote pairs at the single-subject level. Specifically, we used the normalized differences between the estimated learning parameters of two consecutive sessions as contrast weights (partial derivative of performance with respect to time) and entered the resulting contrast images (one image per subject) into a one-sample t test.

For all analyses, the threshold was set to p < 0.05, corrected for multiple comparisons. According to results from previous studies, several regions of interest were defined (medial frontal gyrus, retrosplenial cortex, inferior and superior parietal cortex, parahippocampal gyrus, hippocampus, and caudate nucleus), and correction for multiple comparisons was based on these regions. We applied spherical search volumes ranging from 2056 cm³ (hippocampus) to 4120 cm³ (retrosplenial cortex). Elsewhere in the brain, correction was based on the entire search volume.

Results

Discussion

The present study attempted to identify the neural foundations of successfully forming a survey representation derived from ground-level navigation. We used our recently introduced virtual-reality paradigm, because it provides an objective distinction between route and survey knowledge (Wolbers et al., 2004). Eleven of 17 participants showed clear evidence for survey learning, reflected by increasing accuracy and decreasing response times independent of the varying temporal delays. Neuroimaging results revealed a dissociation between the contributions of the retrosplenial cortex and the hippocampus. Whereas retrosplenial activation paralleled behavioral measures of map expertise, BOLD responses in the left hippocampus indicated the extent of learning in a given session. We therefore conclude that, in the context of navigational learning, the hippocampus mainly serves to incorporate additional information into an emerging memory representation.

Significant learning effects for direct, close, and remote pairs were observed in both behavioral measures. This clearly demonstrates that spatial learning was not confined to the formation of temporospatial associations between consecutive landmark views, as shown during route learning (Wolbers et al., 2004). Instead, participants were able to infer a configurational representation from ground-level navigation, because performance improvements occurred in all retrieval categories, despite the substantially varying temporal delays. The excellent mapdrawing performance lends additional support to this interpretation. However, we cannot exclude the possibility that participants may have acquired some route knowledge as well, which could explain the overall differences revealed by the post hoc tests. Most interestingly, giving specific instructions that point to an allocentric learning strategy seems necessary to induce survey learning, because most participants in our previous study, when only unspecific instructions were given, showed evidence for route learning only (Wolbers et al., 2004).

We can only speculate why six participants failed to meet our criteria for survey learning. Mental abilities, such as spatial visualization and spatial orientation, contribute to interindividual differences in acquiring spatial information from a virtual environment (Waller, 2000). These abilities might be less pronounced in the excluded participants, which makes it impossible to infer a substantial amount of survey knowledge within six exposures to the complex environment. Reliable insights into the neural correlates of survey learning can therefore only be obtained from those participants who showed clear evidence for inferring an allocentric representation.

What processes are needed to transform a virtual environment experienced from a first-person perspective into a survey representation? During navigation, spatial positions of landmarks are perceived in egocentric coordinates. Because the coordinates change with observer movements, optic flow needs to be taken into account to assess current heading direction and the extent of translational movements. This information allows for dynamic spatial updating and for computing displacement vectors (Muehl and Sholl, 2004). Spatial relationships between landmarks can then be estimated, based on an intrinsic reference system that presumably depends on the initial orientation of the observer (Shelton and McNamara, 2004).

Contrasting encoding with the control condition revealed a network of areas that have been repeatedly observed during spatial learning in virtual environments (Aguirre et al., 1996; Shelton and Gabrieli, 2002; Iaria et al., 2003). Whereas the recruitment of the medial frontal gyrus might reflect working memory processes necessary for long-term encoding (Leung et al., 2002; Suzuki et al., 2002), inferior and superior parietal regions process spatial information in egocentric reference frames (Andersen et al., 1997; Galati et al., 2001; Halligan et al., 2003). The lack of significant effects in the parahippocampal cortex may be attributable to its primary role in viewpoint-dependent processing of spatial scenes (Aguirre et al., 1996; Epstein et al., 2003). Given that identical landmarks were presented in a very similar virtual world in the control compared with the encoding condition, parahippocampal involvement would be expected in both conditions. Consequently, these activations should cancel each other out in direct comparisons.

We observed performance-dependent activation increases only in the anterior bank of the parieto-occipital sulcus and in the posterior part of the retrosplenial cortex. In the monkey brain, retrosplenial areas are densely connected with various medial temporal regions, the posterior parietal cortex, and the mid-dorsolateral prefrontal cortex (Maguire, 2001; Kobayashi and Amaral, 2003). Moreover, the retrosplenial cortex of the rat contains headdirection cells that indicate the current heading of the animal (Cho and Sharp, 2001). It is, consequently, in a perfect position to integrate egocentric spatial information from the posterior parietal cortex with information about self-motion and landmark identity originating, at least in part, from medial temporal areas. When this process is disrupted by retrosplenial lesions, severe problems with forming or recalling links between landmark identity and directional information arise (Aguirre and D'Esposito, 1999). Therefore, we believe that the learning-related activation increase observed in bilateral retrosplenial cortices reflects this increasingly robust integration of different sources of spatial information. Constituting a necessary prerequisite for forming survey representations, a more efficient integration allows for faster and more accurate access to the spatial relationships between landmarks during retrieval.

Unlike the performance-dependent retrosplenial activation, BOLD responses in the hippocampus predicted the extent of learning in a given session. In the context of learning a spatial layout, geometric determinants of simple environments (i.e., proximity of local walls) can shape hippocampal place-cell firing in the rat over repeated exposure (Lever et al., 2002). A coherent representation of a complex environment, such as the one used in the present study, however, requires the integration of multiple local environments (i.e., intersections) that are not surrounded by overall boundaries. Such a representation also contains spatial relationships between superordinate entities, such as intersections or regions, which have been shown to substantially influence navigational behavior (Wiener and Mallot, 2003). We therefore suggest that the hippocampus serves to integrate both the layout of local parts and the spatial relationships between superordinate entities into a coherent representation. In early learning stages, the hippocampal contribution is very strong, because the existing representation is only rudimentary, but, as learning progresses, the amount of new associations to be assembled gradually decreases. As a consequence, hippocampal activation becomes weaker in later stages of learning, as has been shown for simple geometric patterns (Grön et al., 2001). This interpretation is supported by the positive correlation between the BOLD response in the hippocampus and the error rate as well as the reaction time that was observed for participants using a spatial strategy during maze learning (Iaria et al., 2003). When these participants made many errors and took a long time to perform the task, hippocampal activation was high, interpreted as indicating the importance of the hippocampus for the learning process.

Activation in the posterior parietal cortex has been shown to parallel behavioral performance during route learning, reflecting the successful construction of spatial links between consecutive landmarks (Wolbers et al., 2004). Although subjects may have acquired some knowledge about the route in the present study as well, the parietal response presumably was attenuated as participants actively focused on inferring survey knowledge. The crucial task of transforming spatial relationships, initially perceived in egocentric coordinates, into an observer-independent reference frame rather depends on retrosplenial and hippocampal functioning (Maguire, 2001; Burgess et al., 2002). Although the importance of the hippocampus for retrieving allocentric representations has been questioned (Teng and Squire, 1999; McNamara and Shelton, 2003; Squire et al., 2004), converging evidence highlights its role in the learning process. This densely connected area may best be described as a modality-independent, additional working memory buffer that serves to associate complex unimodal or multimodal information (Hölscher, 2003). It seems especially important whenever associations need to be formed in a way that allows their flexible use to guide future behavior, as opposed to rigid stimulus-response associations (O'Keefe and Nadel, 1978; Packard et al., 1989; Packard and White, 1990; McNamara and Shelton, 2003). In the current study, high complexity was imposed by the need not only to associate local and superordinate spatial entities but also to transform egocentric coordinates into an observer-independent representation. Forming such a flexible representation may just require the additional recruitment of the hippocampus.

With regard to a possible hemispheric asymmetry, the right hippocampus is often assumed to be especially important for spatial memory, whereas the left hippocampus may be more involved in context-dependent episodic memory (Maguire et al., 1998; Burgess et al., 2002). Although we did not observe significant learning-related changes in the right hippocampus, this does not rule out its potential involvement, because the null hypothesis cannot be verified. However, it is worth noting that there are conflicting findings regarding a hemispheric specialization in the human hippocampus. Several neuroimaging and lesion studies have demonstrated a bilateral or even left-sided hippocampal involvement in spatial memory (Grön et al., 2000; Iaria et al., 2003; Incisa della Rocchetta et al., 2004), which is not in agreement with a strict dissociation of hippocampal functions.

The results of the current study point to different roles of the retrosplenial cortex and the hippocampus during the construction of a survey representation out of ground-level navigation. Having established such a functional dissociation, subsequent studies may now attempt to assess the dynamic interactions between these and other regions that are necessary for the transfer and integration of the described sources of (non)spatial information.