The first step in analyzing behavioral task performance was the calibration of the subject's pupil location (gaze position) to eye position on the projection screen, expressed in cm. A special calibration paradigm was recorded before the beginning of the behavioral tasks, rather than to use the standard EyeLink Calibration performance. Eye position was calibrated using a nine point calibration sequence consisting of four targets at 12 cm eccentricity in the horizontal and vertical direction, four in the corners of a 12 cm × 12 cm square, and one in the center. Each calibration target was presented seven times for 1.2 s. Fixations were included in the calculations when eye velocity remained below 70 deg/s for at least 0.15 s. In each of the nine fixation clusters, fixations with a Mahalanobis distance larger than -2 ln(0.02) were removed as outliers, corresponding to the 98% confidence interval. Following the removal of outliers, the medians of the fixation clusters replaced the individual fixations as calibration input. A 2D polynomial of the 2nd order was used as a calibration function to map the raw horizontal/vertical eye position ṟ_x/ṟ_y onto the calibrated horizontal/vertical eye position c_x/c_y:

Each of the vectors c_x, c_y, ṟx, and ṟ_y in this equation has nine elements, corresponding to the nine fixation positions. The 12 12 parameters (i.e., the two 6D vectors p_x, p_y) of this function were by the least square method on the difference between the calibrated cluster centers centers ([c_x,c_y ]) and the target positions ([ṯ_x, ṯ_y])

where ||A||_F denotes the Frobenius norm of the matrix A. In a hierarchical backward regression, this full polynomial was compared with a reduced polynomial under exclusion of the mixed term

and with a linear calibration function

In 9/13/12 of the 34 subjects the calibration resulted in full polynomial, reduced polynomial/linear functions.

As the learning paradigm lasted longer than 20 min, the calibration functions established during the calibration paradigm were modified by a running offset [o_x, o_y ]

where F_cal (p_x, p_y) denotes the calibration established by the calibration paradigm. The dimension of each of the vectors c_x, c_y, o_x, and o_y is the number of samples recorded during the behavioral task. The running offset was established by a interpolation of grid-points placed at the beginning of every fifth trial. The grid values were fitted to minimize the squared distance between the calibrated fixation locations at the start of each trial and the start location itself. In this way, the calibration corrected for slow head movements during the measurement.

In order to assign visual fixations to their intended targets, as well as avoid the assignment of unrelated explorations, we applied an algorithm previously developed by our working group (Drever et al., 2011b). This algorithm prioritizes the order of the reproduced sequence over the strict spatial accuracy of its constituent target fixations, while also considering omissions, explorations, and order errors. Through the recursive ordered assignment of fixations, it identifies the longest continuous sub-sequence of targets that can be assigned in order, then removes them from the original sequence and restarts, running until no possible assignments remain.

The recall probability was defined for each trial as the number of reproduced and assigned targets divided by the sequence length (20). The initial learning speed was computed as the regression slope of the recall probability expressed as a function of trial number. The final recall probability was defined as the average of the recall probability across the last five trials.

Unsurprisingly, since learning requires attention (Craik et al., 1996), there was a near-complete spatial overlap between the learning and attention responses (such that the learning response was contained within the regions of attention task activation). We were interested in only considering the areas common to both experiments. To this end, we created binary masks for both conditions that corresponded to all regions significant at an FDR threshold of 0.05, along with an additional inclusive mask for gray matter. Only the intersection between all three masks with cluster sizes >100 voxels were considered regions of interest. Seven of such regions of overlapping activity were found throughout the parietal, temporal, and frontal lobes. On each side of the parietal cortex were two adjacent clusters which were combined, resulting in a total of five regions of interest, each defined by a corresponding binary mask. We then calculated the center of mass for each region of interest (Table 1). Mean beta values for each ROI were calculated on the subject level.

Prominent activation of V5 by the spatial learning task serves to confirm its role in the processing of moving visual stimuli (Maunsell and Van Essen, 1983; Moutoussis and Zeki, 2008). Interestingly, however, although this function is considered by some (Wilson et al., 1992) to also be attributable to V1, we did not find activity here when using our learning predictor. This suggests that there may indeed be a specific role performed by this occipito-temporal region during the learning of visual sequences (Fan et al., 2020). For example, it has been suggested that separate pathways exist for the processing of slow vs. fast visual motion, wherein V1 is bypassed in the case of fast movements. In this case, subcortical structures are thought to convey visual information directly from the superior colliculus to V5. A study by Grasso et al. (2018) demonstrated that when V5 activity was transiently disrupted by transcranial magnetic stimulation, subjects performed more poorly on a spatial processing task involving fast (23 deg/sec motion) but not slow (4.4 deg/sec) stimuli. However, this direct colliculo-extrastriate pathway is primarily believed to support implicit processing (Grasso et al., 2018), whereas our paradigm involved explicit awareness of the learning task.

Perhaps unsurprisingly, V1 was not found to be more heavily recruited when targets were in the process of being learned, when contrasting with those either already committed to long-term memory or far ahead enough in the sequence to not yet be attempted. One explanation would be that V1 activity was similarly taking place across the entire experiment, further indicating that it is not being specifically recruited during the processing of targets that are being learned. We did also observe stronger response in V5 than in V1 in the attention task.

Both conditions elicited activity bilaterally in the middle frontal gyrus. The left precentral gyrus was also present in both tasks. As the frontal eye fields are known for their role in directing top-down, goal-directed allocation of visual attention, this is consistent with literature (Crowne, 1983; Cazzoli et al., 2015; Veniero et al., 2021). Mechanisms are believed to involve direct mediation of excitability of lower level visual areas through perceptually-relevant shifts in neural oscillations (Veniero et al., 2021). Additionally, this region also plays an established role in mental rotation/imagery in spatial contexts (Cona and Scarpazza, 2019). However, due to its explicit recruitment during attentional orienting, this function likely contributed to a relatively diminished difference in activation when comparing the learning to attention tasks, further suggesting this region was not specific to learning.

The contribution of the superior parietal lobule (SPL) to visuospatial processing and memory has been demonstrated by numerous studies. A meta-analysis of hundreds of imaging studies on spatial cognition further consolidated the functions of the superior parietal lobule across related cognitive domains, particularly spatial attention, visuospatial working memory, and spatial imagery (Berthoz, 1997; Tagaris et al., 1998; Cona and Scarpazza, 2019). In our study, the superior parietal lobule was bilaterally activated in both the learning and attention tasks, but with greater activation in the learning condition. Of interest is its purported role in spatial imagery, which would be necessary during explicit recall of a visual sequence. As, by definition, this function should be similarly activated in explicit encoding, it may be the case that this parietal activity supports a top-down mechanism reflective of progress in the learning of the sequence.

Applied in a broader context, the SPL has also been shown to be involved during reading, as processing text also requires the ordered incorporation of visual targets into a larger series. This can be illustrated through the phenomenon of simultanagnosia. In this neurological disorder, patients experience an extreme restriction of visuospatial attention, leaving them able to only attend to one item at a time (Wolpert, 1924). The functional components of this condition can be deduced through the presence of lesions in the superior parietal lobule (SPL) (Vialatte et al., 2021), a region where hypoactivity has also been observed in subjects with developmental dyslexia (Peyrin et al., 2011; Reilhac et al., 2013). Its activation during the acquisition of less contextual sequences in our paradigm would consequently also be expected.

In terms of hemispheric differences, literature has indicated a lateralization of function with regards to both working and long-term memory specialization, with the right hemisphere being more heavily recruited during spatial tasks (Smith and Milner, 1981; van Asselen et al., 2006) and the left hemisphere preferentially activating during verbal and object-based memory exercises (Dennis and Whitaker, 1976; Ojemann and Dodrill, 1985). Our results corroborate these conclusions, in that a slight, right-biased asymmetry between hemispheres was present in the activation patterns we observed.

The hippocampus is not only encoding at the moment of presentation, but likely also engaged in the continued consolidation of earlier targets. Research has indicated that long-term consolidation can indeed begin immediately following encoding, as per analysis of changes in default mode network complexity (McDonough et al., 2019). It would therefore be reasonable to assume sustained hippocampal activation throughout the experiment, which could result in a lack of significant activation when contrasting presumed encoding of a specific target with all other activity.

The regions activated by our attentional task largely reflected the components of the dorsal frontoparietal network (Corbetta and Shulman, 2002; Fox et al., 2005; Szczepanski and Saalmann, 2013) including the superior parietal lobule and frontal eye fields. Given that our attention paradigm was explicitly designed to elicit top-down attention to the visuospatial targets, this is as expected (Herbet and Duffau, 2022). Furthermore, when comparing the average activity of our ROIs, we found greater differences between conditions for the superior parietal lobule and V5 than for the MFG (indicated by the significant interaction between region and task). This lends support to the notion that the frontal regions we observed reflect attentional mechanisms not involved with the learning of visuospatial sequences.

When comparing learning task performance (final recall probability, specifically) with beta values in regions of interest as generated using our learning predictor, we did not find a significant correlation. This suggests that the responses we observed in the parietal and temporal ROI, although they were specific to the learning task, were not themselves predictive of the quantity of stored information, nor how efficiently it was stored, but rather marked brain activity related to a learning-specific processing of this visuospatial information.

On the question of whether our learning task was truly explicit in nature, the extensive overlap between learning and attention-only conditions lends support to that conclusion, as implicit learning would not have required attention and a dissociation in functional neuroanatomy would have more likely been present (Vuilleumier et al., 2005). For example, an early fMRI study on implicit visual memory found increased activation in the bilateral lingual gyri in response to the learning of deliberately ignored items; this is distinct from the regions we observed. Furthermore, a Wilkinson et al. (2010) study using transcranial magnetic stimulation found that inhibiting the primary motor cortex abolished implicit learning, showing it to be critical to this form of memory. Activation stemming from our learning paradigm did not include M1, lending additional support to the notion that we may have indeed initiated explicit memory mechanisms. Furthermore, previous findings by our group (Drever et al., 2011a) have shown that fixation accuracy in the DILSS task does not improve across subsequent trials, indicating the absence of implicit learning. That is, the slope of the linear regression of the reproduction error on trial number did not differ significantly from zero, while the mean of this slope across subjects was negative (−0.01 deg/trial). Compared to the error present in the first reproduction (around 2.5 deg), the total improvement in accuracy after 25 trials was only 10%. This absence of improvement therefore indicates that implicit learning plays a subordinate role in this paradigm, even if it cannot be completely ruled out. Likewise, we cannot from our findings conclude that the activated regions play no role in implicit learning. However, the embedding of learning-specific neuronal activity in attention-specific activity in our experiment suggests that both paradigms rely on similar mechanisms, which are likely explicit.

It is important to address the functional overlap in regions identified by the learning predictor, in that they may also be attention-related. However, by contrasting encoding vs. non-encoding states, we were able to identify those that, while also involved with attention, are additionally sensitive to encoding. The SPL (particularly the precuneus), for example, is sensitive to item familiarity (Kim and Cabeza, 2007; Wheeler and Buckner, 2003) and V5 is known to feature priming-sensitive cells (DeAngelis and Newsome, 2004), alongside its role in attention orienting. We therefore do not consider this as confounding, but rather an affirmation of the current understanding of the roles played by V5 and the SPL.

As previously mentioned, there exist potential clinical applications for our findings. Previous research by our group (Eggert et al., 2022) on the encoding of long visuospatial sequences in temporal lobe resection patients demonstrated the usefulness of the DILSS paradigm in detecting impairments in visuospatial memory. Unlike other currently used measures of memory, such as the California Verbal Learning (Elwood, 1995) or Rey-Osterrieth Complex Figure Tests (Shin et al., 2006) (whose application did not indicate memory impairments in the patient sample), the DILSS paradigm was indeed able to differentiate between patients with unilateral temporal lobe resection and healthy controls. The present study contributes a complementary perspective on these prior findings, offering neurological insights into the potential mechanisms utilized by the brain during the performance of our task. For example, the right-biased hemispheric asymmetry in BOLD activation associated with sequence acquisition validates the previous study's finding of poorer performance when subjects had undergone resections in the right temporal lobe.

Unfortunately, due to the demanding nature of the experiment, eight out of 42 scanned subjects were unable to adequately learn the sequence and were therefore ultimately excluded from final analysis. This may have introduced a bias toward faster-learning subjects who found the task easier and possibly acquired the sequence using different neurocognitive resources. As an example, some subjects may have relied more heavily on their superior working memory capacity, rather than encode one target per trial, as was the average rate. This could be resolved in future studies by implementing more stringent screening during the initial behavioral training prior to scanning.

Furthermore, differences between our two tasks carry the possibility of introducing nuisance variables into our analysis. There was, however, a significant interaction effect that indicated the parietal and temporal regions of interest were more responsive to the presence of learning than the MFG was. This cannot plausibly be attributed solely to factors such as target shape or other visual context. For example, targets moved in a similar fashion and although target shape differed, it would be reasonable to assume that our parietal activation was more related to object localization than shape identification. It is also less plausible that the activity in V5 in response to sequence learning demands would be a consequence of such differences in shape. Despite these task differences, their overall contexts were kept as similar as possible. The visual environment was consistent, with white targets against a black background. The distance and angle of target motion during presentation was also similar. Most importantly, the timing of attentional focus was deliberately designed to be analogous, as described in the Methods section. Moreover, a prior study (Gauthier et al., 2002) that also made use of a task involving discrimination between mirrored objects specifically did not find activity in the SPL, as we did. Differences in target shape and character complexity are therefore unlikely to account for the increased activity in the SPL and V5 during the DILSS paradigm. Rather, the specificity of this activation is more likely due to the presence of learning, reflecting associated processes underlying it.

An additional limitation lies in the methodology itself, with 3T fMRI potentially being insufficient for detecting small regions with a low signal-to-noise ratio. Significant activity in the superior colliculus, for example, was not observed. However, its location in the midbrain and proximity to vasculature as well as other sources of physiological noise may have masked its signal. This could potentially be mitigated through the use of a 7T scanner.