Abstract
Our brains are in a constant state of generating predictions, implicitly extracting environmental regularities to support later cognition and behavior, a process known as statistical learning (SL). While prior work investigating the neural basis of SL has focused on the activity of single brain regions in isolation, much less is known about how distributed brain areas coordinate their activity to support such learning. Using fMRI and a classic visual SL task, we investigated changes in whole-brain functional architecture as human female and male participants implicitly learned to associate pairs of images, and later, when predictions generated from learning were violated. By projecting individuals’ patterns of cortical and subcortical functional connectivity onto a low-dimensional manifold space, we found that SL was associated with changes along a single neural dimension describing covariance across the visual-parietal and perirhinal cortex (PRC). During learning, we found regions within the visual cortex expanded along this dimension, reflecting their decreased communication with other networks, whereas regions within the dorsal attention network (DAN) contracted, reflecting their increased connectivity with higher-order cortex. Notably, when SL was interrupted, we found the PRC and entorhinal cortex, which did not initially show learning-related effects, now contracted along this dimension, reflecting their increased connectivity with the default mode and DAN, and decreased covariance with visual cortex. While prior research has linked SL to either broad cortical or medial temporal lobe changes, our findings suggest an integrative view, whereby cortical regions reorganize during association formation, while medial temporal lobe regions respond to their violation.
Significance Statement
The current work is the first to investigate changes in whole-brain manifold architecture that underlie visual statistical learning (SL). We found that areas of the visual cortex and dorsal attention network showed significant connectivity changes during learning, reflecting their decreased, and increased covariance with other networks, respectively. Notably, when SL was later disrupted, regions within the medial temporal lobe, which had shown no evidence of initial learning, now began to increase connectivity with higher-order cortex. Together, these findings not only reveal the widespread neural interactions that underlie visual SL but also extend prior work, suggesting separable cortical and medial temporal lobe contributions for the encoding versus violation of learned associations.
Introduction
Each day, our central nervous systems are inundated with a constant influx of sensory information. Despite the apparent randomness of these inputs, we possess the remarkable ability to discern these repeating patterns in our environment. Our ability to implicitly extract regularities over time and across development is known as statistical learning (SL), and it enables us to make predictions about the world and modify our actions accordingly (Saffran, 2009; Turk-Browne, 2012; Aslin, 2017; Forest et al., 2023). One particularly well-studied form of such learning is visual SL, which examines our ability to generate internal representations of, and extract probabilities for, visual stimuli that co-occur in time (Turk-Browne et al., 2005, 2008).
Traditionally, SL was thought to engage only task-specific sensory cortices, while episodic learning was associated with the medial temporal lobe (MTL), primarily the hippocampus, along with the entorhinal, parahippocampal, and perirhinal cortex (PRC; McClelland et al., 1995; O’Reilly et al., 2014). However, recent evidence has challenged this neural distinction. SL has been shown to involve not only sensory-specific cortices but also higher-order association networks (Huettel et al., 2002; Karuza et al., 2017; Karlaftis et al., 2019), suggesting that it requires both task-specific processing (e.g., visual regions; Turk-Browne et al., 2009, 2010) and task-general cognitive functions like attention shifting and resource allocation (Saffran and Thiessen, 2007; Frost et al., 2015, 2019). Additionally, contrary to previous beliefs that regularities were learned exclusively within cortical networks, studies now show that the MTL and hippocampus exhibit representational change and differential activity during SL (Schapiro et al., 2012, 2016; Wammes et al., 2022). Consequently, emerging views suggest that both cortical and subcortical regions contribute to SL, yet how these regions coordinate their activity remains unclear. At the very least, it seems likely that stimulus-relevant sensory cortical regions must share information with higher-order systems to allow learning to take place.
One approach to examining this problem has come from the investigation of changes in functional connectivity between individual pairs of brain regions associated with SL (Karuza et al., 2017; Park et al., 2022). While such studies have highlighted connectivity changes across several regions, the importance of any individual connection is ambiguous given that any single region projects widely across several areas and brain networks (Bullmore and Sporns, 2009; Van Den Heuvel et al., 2009). Indeed, while prior work points to the involvement of the MTL, hippocampus, as well as task-specific and task-general cortical networks in SL, these brain areas project widely across the cortex and are reciprocally interconnected with several, distributed whole-brain networks (Froudarakis et al., 2019; Barnett et al., 2021). It stands to reason that inter-regional couplings might play a role in the learning process. Moreover, as learning involves the integration of information processing across multiple brain areas (Froudarakis et al., 2019; Barnett et al., 2021), understanding the importance of observed connectivity changes requires consideration of the entire whole-brain networks in which each brain area is embedded. This is necessary for resolving questions about the extent and timing of connectivity changes during SL and how different brain systems may independently or cooperatively encode statistical regularities (Batterink et al., 2019; Marlatte et al., 2024).
Recent advances in dimension reduction techniques, adapted from electrophysiology (Shenoy et al., 2013; Cunningham and Yu, 2014), have provided new tools for studying large-scale patterns of brain activity in functional magnetic resonance imaging (fMRI; Margulies et al., 2016; Huntenburg et al., 2018; Paquola et al., 2019; Gale et al., 2022; Nick et al., 2024). These methods allow high-dimensional connectivity patterns to be characterized along a low-dimensional subspace or manifold, reflecting the major patterns of covariance across brain areas. This approach has revealed new insights into the organization of large-scale neural activity in both healthy and clinical populations (Shenoy et al., 2013; Hong et al., 2019; Vyas et al., 2020; Xia et al., 2022). More recent studies have extended this technique to investigate how whole-brain manifold structures evolve during learning tasks, with results implicating networks that had been previously overlooked when explaining learning mechanisms (Gale et al., 2022; Song et al., 2023; Nick et al., 2024). In this study, we applied this manifold learning approach to visual SL for the first time, allowing us to examine how whole-brain cortical and MTL activity patterns are coordinated during different stages of learning. This exploratory analytic approach allowed us to provide a first-of-its-kind characterization of the dynamic landscape of brain activity during SL and better understand how brain regions collectively adapt to changing environmental patterns. Given that prior connectivity studies suggest that both attention networks and MTL structures contribute to SL, we reasoned that our whole-brain dimension reduction approach might detect the emergence of these potentially interacting networks.
Materials and Methods
Imaging data were preprocessed using fMRIPrep, which is open source and freely available. Our analysis and links to data are available at https://github.com/thelamplab/visualSLmanifolds.
Participants
Our study performed a reanalysis on data collected from 33 healthy individuals (21 female) between the ages of 18–35 years old (Wammes et al., 2022). All participants reported having normal visual acuity (either uncorrected to corrected to normal) and good color vision. All participants gave informed consent to a protocol approved by the Yale Institutional Review Board and received $20 per hour as compensation for their time. Eight participants were excluded from the original data's 41 participants due to either missing functional runs (n = 7) or issues in the quality of hippocampal and cortical segmentations in the present analysis (n = 1).
Statistical learning task
To investigate changes in whole-brain functional connectivity during SL, we analyzed human fMRI data collected during a classic visual SL task, where abstract images were presented to subjects one at a time on screen (Wammes et al., 2022). The task stimuli included eight unique image pairs, generated using a neural network to ensure that they vary in similarity from having no discernable shared features to being nearly indistinguishable from each other (Fig. 2A). The task was composed of eight functional runs, each lasting 304.5 s (203 TRs), where images were presented for 1 s each. These functional runs took two different forms: structured or random. In the structured runs, unbeknownst to the participant, the sequence of images presented to the participant had the eight image pairs (e.g., AB) embedded with their paired structure preserved, such that when a given image (e.g., A) was presented, it was always followed by its pairmate (e.g., B). In these structured runs, a given pair was never presented twice in a row.
The six structured runs were book-ended by two random runs, where images were presented in an unstructured order (i.e., unpaired), thereby allowing us to establish the characteristic patterns of whole-brain connectivity that manifest prior to learning (Pre-learning run), as well as after learning had occurred, in a situation where the paired sequence structure was interrupted (Post-learning run). During both structured and random runs, a small, partially transparent gray patch was overlaid on 10% of images (Fig. 2B). The participants were instructed to perform a cover task of pressing a button on a handheld box when they noticed the gray patch. This secondary task was entirely orthogonal to SL, serving only to maintain attention on the images throughout the task.
MRI acquisition and preprocessing
All data collected from prior research (Wammes et al., 2022) was scanned using a 3T Siemens Prisma scanner with a 64-channel head coil at the Yale Magnetic Resonance Research Center. For each participant, eight functional runs were collected with a multiband echoplanar imaging (EPI) sequence (TR, 1,500 ms; TE, 3.26 ms; voxel size, 1.5 mm isotropic; FA, 71; multiband factor, 6), yielding 90 axial slices. Each run contained 203 volumes. Here, our results come from preprocessing performed using fMRIPrep 20.1.2 (Esteban et al., 2019, 2024; RRID: SCR_016216), which is based on Nipype 1.5.1 (Gorgolewski et al., 2011; Esteban et al., 2022; RRID: SCR_002502). Many internal operations of fMRIPrep use Nilearn 0.6.2 (Abraham et al., 2014, RRID:SCR_001362), mostly within the functional processing workflow. For more details of the pipeline, see the section corresponding to workflows in fMRIPrep's documentation.
Anatomical data preprocessing
A total of one T1-weighted (T1w) image was found within the input BIDS dataset. The T1-weighted (T1w) image was corrected for intensity nonuniformity (INU) with N4BiasFieldCorrection (Tustison et al., 2010), distributed with ANTs 2.2.0 (Avants et al., 2008, RRID:SCR_004757), and used as T1w reference throughout the workflow. The T1w reference was then skull stripped with a Nipype implementation of the antsBrainExtraction.sh workflow (from ANTs), using OASIS30ANTs as target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM) and gray matter (GM) was performed on the brain-extracted T1w using fast (FSL 5.0.9, RRID:SCR_002823, Zhang, et al., 2001). Brain surfaces were reconstructed using recon-all (FreeSurfer 6.0.1, RRID:SCR_001847 (Dale et al., 1999), and the brain mask estimated previously was refined with a custom variation of the method to reconcile ANTs-derived and FreeSurfer-derived segmentations of the cortical gray matter of Mindboggle (RRID:SCR_002438, Klein et al., 2017). Volume-based spatial normalization to two standard spaces (MNI152NLin6Asym, MNI152NLin2009cAsym) was performed through nonlinear registration with antsRegistration (ANTs 2.2.0), using brain-extracted versions of both T1w reference and the T1w template. The following templates were selected for spatial normalization: FSL's MNI ICBM 152 nonlinear 6th Generation Asymmetric Average Brain Stereotaxic Registration Model (Evans et al., 2012; RRID:SCR_002823; TemplateFlow ID: MNI152NLin6Asym) and ICBM 152 Nonlinear Asymmetrical template version 2009c (Fonov et al., 2009; RRID:SCR_008796; TemplateFlow ID: MNI152NLin2009cAsym).
Functional data preprocessing
For each of the eight BOLD runs found per subject (across all tasks and sessions), the following preprocessing was performed. First, a reference volume and its skull stripped version were generated using a custom methodology of fMRIPrep. Head-motion parameters with respect to the BOLD reference (transformation matrices and six corresponding rotation and translation parameters) are estimated before any spatiotemporal filtering using mcflirt (FSL 5.0.9; Jenkinson et al., 2002). BOLD runs were slice time corrected using 3dTshift from AFNI 20160207 (Cox and Hyde, 1997, RRID:SCR_005927). A B0-nonuniformity map (or fieldmap) was estimated based on two (or more) EPI references with opposing phase-encoding directions, with 3dQwarp (Cox and Hyde, 1997; AFNI 20160207). Based on the estimated susceptibility distortion, a corrected EPI reference was calculated for a more accurate coregistration with the anatomical reference. The BOLD reference was then coregistered to the T1w reference using bbregister (FreeSurfer) which implements boundary-based registration (Greve and Fischl, 2009). Coregistration was configured with 6 degrees of freedom. The BOLD time series were resampled onto the following surfaces (FreeSurfer reconstruction nomenclature): fsaverage5, fsaverage. The BOLD time series (including slice-timing correction when applied) were resampled onto their original, native space by applying a single, composite transform to correct for head-motion and susceptibility distortions. These resampled BOLD time series will be referred to as preprocessed BOLD in original space or just preprocessed BOLD. The BOLD time series were resampled into standard space, generating a preprocessed BOLD run in MNI152NLin6Asym space. First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. Grayordinates files (Glasser et al., 2013) containing 91,000 samples were also generated using the highest-resolution fsaverage as intermediate standardized surface space. Automatic removal of motion artifacts using independent component analysis (ICA-AROMA; Pruim et al., 2015) was performed on the preprocessed BOLD on MNI space time series after removal of non-steady-state volumes and spatial smoothing with an isotropic, Gaussian kernel of 6 mm FWHM (full-width at half-maximum). Corresponding “nonaggresively” denoised runs were produced after such smoothing. Additionally, the “aggressive” noise regressors were collected and placed in the corresponding confounds file. Several confounding time series were calculated based on the preprocessed BOLD: framewise displacement (FD), DVARS, and three region-wise global signals. FD was computed using two formulations following Power (absolute sum of relative motions; Power et al., 2014) and Jenkinson (relative root mean square displacement between affines; Jenkinson et al., 2002). FD and DVARS are calculated for each functional run, both using their implementations in Nipype (following the definitions by Power et al., 2014). The three global signals are extracted within the CSF, the WM, and the whole-brain masks.
Additionally, a set of physiological regressors were extracted to allow for component-based noise correction (CompCor; Behzadi et al., 2007). Principal components are estimated after high-pass filtering the preprocessed BOLD time series (using a discrete cosine filter with 128 s cutoff) for the two CompCor variants: temporal (tCompCor) and anatomical (aCompCor). tCompCor components are then calculated from the top 5% variable voxels within a mask covering the subcortical regions. This subcortical mask is obtained by heavily eroding the brain mask, which ensures it does not include cortical GM regions. For aCompCor, components are calculated within the intersection of the aforementioned mask and the union of CSF and WM masks calculated in T1w space, after their projection to the native space of each functional run (using the inverse BOLD-to-T1w transformation). Components are also calculated separately within the WM and CSF masks. For each CompCor decomposition, the k components with the largest singular values are retained, such that the retained components’ time series are sufficient to explain 50% of variance across the nuisance mask (CSF, WM, combined, or temporal). The remaining components are dropped from consideration. The head-motion estimates calculated in the correction step were also placed within the corresponding confounds file. The confound time series derived from head-motion estimates and global signals were expanded with the inclusion of temporal derivatives and quadratic terms for each (Satterthwaite et al., 2013). Frames that exceeded a threshold of 0.5 mm FD or 1.5 standardized DVARS were annotated as motion outliers. All resamplings can be performed with a single interpolation step by composing all the pertinent transformations (i.e., head-motion transform matrices, susceptibility distortion correction when available, and coregistrations to anatomical and output spaces). Gridded (volumetric) resamplings were performed using antsApplyTransforms (ANTs), configured with Lanczos interpolation to minimize the smoothing effects of other kernels (Lanczos, 1964). Nongridded (surface) resamplings were performed using mri_vol2surf (FreeSurfer).
Investigating neural changes in the hippocampus in response to statistical learning
Evidence supporting visual SL in this task comes from a reanalysis of the original paper reporting on this dataset (Wammes et al., 2022), confirming that the critical learning effects remain when including only the participants we were able to include in our analyses. Here, we tested for evidence of representational change in line with predictions from the non-monotonic plasticity hypothesis (NMPH; Detre et al., 2013; Ritvo et al., 2019, 2023; Midler and McClelland, 2023). That is, representational change is expected to follow a cubic function based on image pair similarity, where dissimilar pairmates will show no representational change and highly similar pairmates show evidence of neural integration, while moderately similar pairmates will show evidence of neural differentiation.
As a summary of the representational change analysis used in the original paper (see Wammes et al., 2022 for a full overview), a general linear model was fit for each of the 16 images for the two random runs (here, referred to as Pre- and Post-learning epochs). The parameter estimates corresponding to each unique image were extracted from the voxels within each region of interest (CA1, CA2/3, DG, and whole hippocampus) and vectorized. For each image pair, we computed Pearson’s correlation between the two vectors representing the brain activity corresponding to each image in the pair, yielding eight distinct representational similarity values for both Pre- and Post-learning epochs. These Pre-learning values were subtracted from Post-learning scores to examine how the representations of each pair's images changed with respect to one another over the course of visual SL (i.e., from Early- to Late-learning). Here, positive scores indicate integration, while negative scores relate to differentiation between representations of two images in a given pair. To test whether these representational change scores were consistent with learning-related change predicted by the NMPH, we fit a theory-constrained cubic model to the data from all but one held-out participant and tested how well this model predicted the held-out participant's data. This analysis was completed for each participant, and the resulting correlations between each participant's observed and model-predicted values were submitted to bootstrap-resampled t tests.
Cortical and subcortical region time series extraction
For each participant and each of our task epochs (Pre-, Early-, Late-, and Post-learning), we extracted the mean blood oxygen level-dependent (BOLD) time series data for each of the 998 cortical regions predefined by the Schaefer 1,000 parcellation (Schaefer et al., 2018) and each of the 14 subcortical regions of interest (MTL regions and hippocampal subfields: left and right CA1, CA2/3, DG, subiculum, ERC, PRC, and PHC) predefined using the individual-specific parcellations derived from automated segmentations of hippocampal subfields (ASHS) software (Wisse et al., 2016) with the Princeton Atlas (Aly and Turk-Browne, 2015). Two cortical regions in the Schaefer 1000 parcellation were removed due to their small parcel size. Our time series data was denoised using a combination of our confound regressors from our initial preprocessing using fMRIprep, alongside discrete cosine regressors (with a high-pass filter threshold of 128 s) generated by fMRIprep. Furthermore, a low-pass filtering process using a Butterworth filter with a cutoff point at 100 s was also applied and implemented using Nilearn. Finally, our cortical and subcortical time series data were z-scored within each region and combined to create a comprehensive time series dataset, including cortical and MTL activity.
Functional connectivity estimation and centering
For each participant, we extracted regional time series data for four equal-length task epochs (each containing 203 imaging volumes). The Pre-learning run (i.e., the first random run), where images were presented to participants in an unstructured order, was used as a baseline. The Early- and Late-learning runs consisted of the first and last unstructured runs in the task, where images were presented in an order that contained an embedded pair structure, such that a given image was always followed by its pairmate. The first and last structured runs were chosen in order to capture a timepoint very early and a timepoint very late in implicit learning. Finally, we used the final random run as our measure of Post-learning. For each of our task epochs (Pre-, Early-, Late-, and Post-learning), we generated functional connectivity matrices by calculating the region-wise covariance matrix using the Ledoit–Wolf estimator (Ledoit and Wolf, 2003).
As demonstrated in prior work (Areshenkoff et al., 2022; Gale et al., 2022), the influence of subject-level clustering on task-based changes in functional connectivity can hinder the analysis of group-level learning-related changes in connectivity. To uncover differences in task-related structure and remove subject-level differences, we used a technique described in prior work (Gratton et al., 2018; Zhao et al., 2018; Areshenkoff et al., 2021) to center the connectivity matrices. This approach leverages the natural Riemannian geometry of the covariance matrix space and makes adjustments to each matrix for each participant, so they all share a common mean, allowing us to eliminate the variations in functional connectivity that are specific to individual participants (Fig. 1B; for details of this centering approach, as well as its benefits, see Zhao et al., 2018).
Overview of functional connectivity analysis and centering approach. A, Functional connectivity analysis approach. For each participant and task epoch (Pre-, Early-, Late-, and Post-learning), we computed functional connectivity matrices from parcel-wise time series data extracted from the Schaefer 1000 cortical parcellation (Schaefer et al., 2018) and participant-specific hippocampal parcellations derived from ASHS (Wisse et al., 2016) software. We then estimated functional connectivity manifolds for each task epoch using PCA with centered and thresholded connectivity matrices (see Materials and Methods). All manifolds were aligned to a common template manifold created from the group-average Pre-learning connectivity matrix using Procrustes alignment. B, Uniform manifold approximation (UMAP; McInnes et al., 2018) visualization of connectivity matrices, both before and after centering. Single data points represent individual matrices. For both uncentered and centered data, the left plots have the data points colored according to subject identity whereas the right plots have the data points coloured according to task epoch. Note that prior to centering, matrices were mainly clustered according to subject identity whereas this cluster structure was abolished following centering.
Dimension reduction and manifold construction
The following steps were used to derive connectivity manifolds from our previously centered functional connectivity matrices. First, in line with prior research (Margulies et al., 2016; Hong et al., 2019; Gale et al., 2022), we applied row-wise thresholding to retain the top 10% of connections in each row of our matrices and remove any weak or spurious connections (see Extended Data Fig. 5-1 for results demonstrating our findings are robust to different thresholding parameters). Next, to characterize the similarity between connectivity profiles of different brain regions, we computed cosine similarity between rows in our matrices to create an affinity matrix. We then used principal components analysis (PCA) to extract a set of principal components (PCs) that provide a low-dimensional representation of the connectivity structure of our data. As a method for dimensionality reduction, PCA simplifies high-dimensional patterns of whole-brain functional connectivity into a low-dimensional space, highlighting the most prominent patterns of covariance across brain networks in fMRI data. Unlike methods that focus on pairwise connections, PCA identifies dominant modes of covariation across multiple regions, revealing system-level network reconfigurations. By focusing on these principal components, PCA offers a more concise and interpretable approach to studying large-scale patterns of brain activity across learning, revealing dynamic network interactions that might be obscured by more granular analyses.
To investigate changes in participants’ functional network architectures across visual SL, we constructed a template manifold from a group-averaged Pre-learning connectivity matrix. This template was calculated from the geometric mean across all participants’ centered Pre-learning connectivity matrices. Subsequently, all individual manifolds (33 participants by four epochs; 132 total manifolds) were aligned to this template manifold using Procrustes transformation (Fig. 1A). This approach allowed us to examine learning-specific deviations away from the template manifold throughout the task in a common low-dimensional neural manifold space, thus enhancing the sensitivity of our analyses.
Word-cloud analysis
The loadings from our top three PCs were interpreted using the Neurosynth Decoder (Yarkoni et al., 2011). This tool employs an automatic meta-analysis and text-mining technique to take specific activations found in different whole-brain maps (in our case, the map for each PC loading; Fig. 3A) to identify keywords that are commonly found in neuroimaging studies with similar maps. It uses this method to generate a collection of keywords that are associated with a given brain mask, along with their correlation values. Our analysis identified the 10 largest positive and negative correlations for each PC's map. We then transformed these lists into word clouds, where the size and color of each term reflected its correlation's direction and strength (Fig. 3E). Importantly, we omitted all terms related to neuroanatomy (e.g., “prefrontal”) and any repeated terms (e.g., “object” and “objects”) from our final selection.
Examining task-related variability in component loadings across learning epochs
Guided by prior work (Hong et al., 2019), we investigated task-related variation within our top 3 PCs by quantifying the variability of each subject's PC loadings across the four task epochs (Pre-, Early-, Late-, and Post-learning). This approach aimed to characterize the range of whole-brain connectivity changes throughout the SL task and identify which PCs, if any, exhibited significant shifts across these epochs of interest. For each PC, we compared the loadings for each region along this axis and found the minimum and maximum loading for each participant. We then subtracted the minimum from the maximum to derive a measure of the min–max range for each participant within each of our task epochs. With these difference score data (i.e., one min–max score for each participant and epoch), we performed a repeated-measures analysis of variance (rmANOVA) across task epochs for each PC. We corrected each rmANOVA for multiple comparisons using false discovery rate (FDR) correction at q < 0.05 (Benjamini and Yekutieli, 2005). The results of each of these rmANOVAs revealed which neural dimensions were most affected as a function of task epoch, allowing us to discern which dimension warranted post hoc comparisons to further investigate changes in regional brain activity. To preface the results, this analysis indicated that only PC1, a component that meta-analysis revealed was implicated in visual processing and object recognition showed significant variation across epochs. From here on, we refer to PC1 as the visual-objects axis.
Examining changes along the visual-objects axis during statistical learning
In our main analysis, we compared each region's loadings on the visual-objects axis (i.e., PC1) during Pre-, Early-, Late-, and Post-learning epochs by performing an rmANOVA for each of the 1,012 regions in our subcortical-cortical manifold. For regions where the rmANOVAs showed significant change in their position along the visual-objects axis, we also performed post hoc paired t tests contrasting pairs of epochs. These additional t tests served to further investigate the specific differences from Pre- to Early-learning, Early- to Late-learning, and Late- to Post-learning. Both the rmANOVAs and t tests were corrected for multiple comparisons using FDR correction (q < 0.05; Benjamini and Yekutieli, 2005).
Exploring regional connectivity underlying changes along the visual-objects axis
To explore the underlying changes in functional connectivity that lead to alterations along the visual-objects axis (PC1), we performed seed connectivity contrasts between different task epochs. To this end, we selected seed regions whose loadings on the visual-objects axis were modulated based on task epoch. Together, the patterns of connectivity changes associated with these regions help to describe the underlying shifts in connectivity that result in the key changes in loadings along the axis that were observed in our main analysis.
For each seed region, we created functional connectivity maps for each participant and epoch of interest and computed region-wise paired t tests for both Late-learning > Early-learning and Post-learning > Late-learning comparisons. For each comparison, we opted to show the unthresholded t-maps, to better visualize and compare the changes in connectivity that drive the observed changes along the visual-objects axis. In addition to these t-maps, we constructed polar plots depicting the specific changes in t-values at the network level, by averaging the correlation values for all areas within each network and then averaging across participants (Yeo et al., 2011). Note that these analyses are mainly intended to provide a visual characterization and interpretation of the connectivity changes of each seed region from our main analysis.
Investigating the role of the hippocampus in visual statistical learning
In order to characterize the specific changes in functional connectivity between the hippocampus and cortex during our task, we performed an additional analysis, contrasting left and right hippocampal seed connectivity maps across our task epochs. We created these left and right hippocampal seed maps by computing functional connectivity maps for bilateral CA1, CA2/3, DG, and subiculum seed regions and then averaging the correlation values from each of the connectivity maps within the set of left and right subregions, respectively. By using hippocampal-based maps, we aimed to examine how the hippocampus specifically interacts with other cortical areas throughout different task epochs, a feature that is predicted from the previous literature (Hindy et al., 2016; Barron et al., 2020), but was not apparent from our main analysis. Similar to our main seed connectivity analysis, we also constructed polar plots for each hippocampal seed region to better depict the specific changes in correlations at the network level.
Results
Statistical learning results in neural changes at the regional level
Prior work, including our own, has shown that visual SL leads to changes in stimulus-specific patterns of neural activity in the hippocampus (Schapiro et al., 2012, 2013, 2016; Wammes et al., 2022). Specifically, depending on the visual similarity between image pairs, learning can lead to either neural integration of representations, where two temporally coupled stimuli become more similar in their neural representations, or neural differentiation, whereby they become less similar. Recent theoretical (Ritvo et al., 2019; Frank et al., 2020), modeling (Midler and McClelland, 2023; Ritvo et al., 2023), and empirical (Detre et al., 2013; Lewis-Peacock and Norman, 2014; Wammes et al., 2022) work indicates that this relationship follows a U-shaped function, as described by the NMPH. In our prior work based on these data (Wammes et al., 2022), the strength of this U-shaped function in the DG provided the primary measure that learning had taken place. Consistent with this prior finding, in our smaller subset of the original data, we observed a U-shaped learning effect in the CA1 (r = 0.11, 95% CI [−0.010, 0.232], p = 0.035) and DG (r = 0.11, 95% CI [−0.017, 0.248], p = 0.045), but not CA2/3 (r = 0.07, 95% CI [−0.044, 0.180], p = 0.115) or across the entire hippocampus (r = 0.07, 95% CI [−0.032, 0.180], p = 0.088; Fig. 2C). These results demonstrate that the existing prior effects remain when using our subsample of the previously published dataset and demonstrate that SL induced the hypothesized representational changes in hippocampus. Building from this, we next examined whether, and to what extent, functional interactions within and outside of the MTL are also altered during SL.
Task structure and representational change during visual SL. A, Image pairs used in the visual SL task (Wammes et al., 2022), ordered from most visually dissimilar, to most visually similar. B, Visual SL task structure. Participants viewed images presented one at a time on screen. Structured runs contained an embedded pair structure, whereby an A image was always followed by its pairmate (B image; example of repeating pair mates denoted by red dashed line). In contrast, random runs presented images in a completely random order. Each run consisted of 80 image presentations (5 repetitions of each image or pair). Participants were tasked with pressing a button when a small faded gray square appeared in the image (10% of trials). C, Left, Relationship between image similarity and representational change predicted by the NMPH. For each region (CA1, CA2/3, DG and hippocampus, HIPP), we correlated parameter estimates derived from a general linear model for each A and B image in a pair and then created comparison scores by subtracting Pre- from Post-learning correlation scores. This measure quantifies the degree of representational change that takes place between image pairs as participants underwent visual SL (i.e., from Early- to Late-learning), with positive scores indicating integration between two images in a pair, and negative scores indicating differentiation. The comparison scores were plotted as a function of similarity level (i.e., coactivation) for each of the eight image pairs (ordered from least to greatest visual similarity in paired items). Shaded areas depict the bootstrap-resampled standard error of the mean (SEMs). Right, Theory-constrained cubic models were fit to the representational change patterns in all but one held-out participant and then used to predict the held-out participants’ scores. These individual-specific model fits are shown in the light gray scatterplots. Bar plots display the group-level model fit, as measured by the correlation between actual and predicted values. At the group-level, representational change was reliably predicted by a theory-constrained U-shaped model in the CA1 and DG. Error bars depict SEM.
Cortical and subcortical manifold structure prior to statistical learning
Prior work indicates that individual differences account for the majority of variance in patterns of brain activity, often obscuring task-related changes in functional connectivity (Gratton et al., 2018; Zhao et al., 2018; Areshenkoff et al., 2022). We recently developed a Riemannian manifold centering approach to mitigate these effects, making task-related changes in brain structure more readily detectable (Areshenkoff et al., 2021; see Materials and Methods). To illustrate the impact of this method, we projected participants’ covariance matrices both before and after this centering procedure using uniform manifold approximation (UMAP; McInnes et al., 2018). Consistent with prior work, including our own (Zhao et al., 2018; Gale et al., 2022; Nick et al., 2024), before centering, we found that the covariance matrices were mainly clustered according to subject identity (Fig. 1B, left). However, after centering, this subject-level clustering was completely abolished, thus permitting exploration of task-related changes in connectivity across epochs (Fig. 1B, right).
Following our centering procedure (see Materials and Methods) to remove subject-level neural structure in our data, the top 3 PCs of our Pre-learning template manifold effectively describe the main dimensions of cortical and subcortical functional connectivity during the task (Fig. 3A), as well as explain the majority of the variance in the data (60.0%; Fig. 3B). PC1 distinguishes the visual network, areas of the parietal DAN, and PRC (positive loadings in red) from other cortical regions (negative loadings in blue). PC2 separates transmodal, default mode network (DMN) regions (in red) from the somatomotor regions (in blue). Lastly, PC3 primarily distinguishes medial areas of the visual cortex (in red) and areas within the DAN (in blue). Together, PC1 and PC2 mainly differentiate visual and somatomotor regions from transmodal regions of the DMN (Fig. 3C, left and middle), approximating the brain's intrinsic functional architecture, as observed in prior work using resting-state data (Margulies et al., 2016; Huntenburg et al., 2018). This tripartite architecture is thought to reflect a crucial feature of cortical brain organization, with the shift from unimodal to transmodal regions of the brain signifying a global processing hierarchy that ascends from lower-order sensory and motor systems to higher-order association areas of the cortex, such as the DMN (Smallwood et al., 2021). For a complete visualization of this tripartite structure, we projected each brain region's loadings on all three PCs onto a three-dimensional manifold space (Fig. 3D).
Functional organization of the Pre-learning epoch. A, Cortical-subcortical loadings for each of the top three PCs, projected in surface space. B, Percent variance explained for the first 20 PCs. C, Histograms displaying the distributions of each network's loading along each PC, colored according to its functional network assignment, denoted on the bottom (Yeo et al., 2011). D, Functional organization of the Pre-learning epoch. Each brain region is depicted as a point in a three-dimensional manifold space, with its location determined by its loading on the first three PCs. E, Meta-analysis of PC loadings submitted to the Neurosynth Decoder tool (Yarkoni et al., 2011). Word clouds show the top and bottom ten functions likely associated with each PC. The text size and color for each word cloud describes their correlation (bigger words have higher correlation values and smaller words have smaller values) and polarity (words in red are more positive and in blue are more negative).
Changes in group-average manifold loadings along the visual-objects axis. A, Changes in functional organization during Early-, Late-, and Post-learning epochs. Each brain region is shown as a point in a two-dimensional manifold space along the visual-objects (PC1) and autobiographical-episodic (PC2) axis, with its color determined by its functional network assignment, denoted at the bottom (Yeo et al., 2011). The horizontal dashed line and bars across epochs (in gray) allows for a direct visual comparison of the change in the expression of the visual-objects axis across task epochs. Histograms along the y-axis show point densities for each epoch and functional network. B, Temporal trajectories represent left and right hemisphere seed regions. The starting position of each seed region from the preceding epoch is marked by an open white circle, with arrows indicating the displacement of each region in the current epoch (colored-in endpoint; Yeo et al., 2011). Gray histograms along the y-axis display the mean density of points across all regions. Inset bar plots within each task epoch display the average change (Δ) in loading for each pair of seed regions from the preceding to current epoch. Negative and positive values reflect decreases and increases along the visual-objects axis, respectively. Single data points depict single subjects. Error bars depict standard error of the mean.
To better understand the typical functions associated with each of our PCs, we conducted a meta-analysis using the Neurosynth Decoder (Yarkoni et al., 2011). This tool quantifies patterns of whole-brain activity against a comprehensive dataset of thousands of fMRI studies. The results of this analysis come in the form of correlation values, revealing each PC's associations with keywords, often indicating specific cognitive functions. For the purposes of visualization, we have transformed these correlation values into word clouds (Fig. 3E) and assigned functional labels to each PC based on these word cloud visualizations. We have termed PC1 the visual-objects axis, emphasizing the role of its constituent regions in object recognition and visual processing; PC2 the autobiographical-episodic axis, given the role of its constituent regions in memory and time-related cognition; and PC3 the working-tasks axis, given the role of its constituent regions in multiple tasks.
While together, the top 3 PCs collectively recreate the tripartite structure—transition from unimodal to transmodal cortex—seen in prior work (Margulies et al., 2016; Huntenburg et al., 2018; Hong et al., 2019; Areshenkoff et al., 2022; Fig. 3D), the order of variance explained by these orthogonal dimensions offers insights into the specific neural demands of our task. Notably, prior resting-state studies using the same dimension reduction method (PCA) have identified our visual-object connectivity axis (PC1) as explaining the third-highest proportion of variance, while our PC2 and PC3 commonly explain the first and second highest variances, respectively (Margulies et al., 2016; Paquola et al., 2019; Gale et al., 2022). However, considering that the stimuli in our task were originally designed to engage higher-order object-selective visual regions (Wammes et al., 2022), it is intuitive that PC1 explains the greatest amount of variance in our data. This shift in variance in our observed task-based PCs, compared with previously reported resting-state gradients (Margulies et al., 2016), highlights how task demands can alter the order and prominence across PCs. Given this, we next examined whether visual SL selectively modulated connectivity motifs along this visual-object axis.
Changes along the visual-object connectivity axis during statistical learning
Figure 4A shows the average loading of each brain region along the visual-object axis as a function of task epoch (Early-, Late-, and Post-learning). Based on prior work (Gale et al., 2022; Nick et al., 2024), to track how each region's connectivity profile shifts across task epochs, we define contraction as a shift toward the component's center (or 0 point) and expansion as a shift away from the component's center. Accordingly, regional shifts along the visual-objects axis, either contractions or expansions, signify changes in a given region's overall connectivity profile. Notably, Late-learning appears to be associated with a contraction along the visual-object axis, followed by an expansion along its top edge once the learned association was interrupted in Post-learning. To direct our focus for post hoc analyses, we tested whether these patterns of contraction and expansion along the visual-objects axis (or any other) was significant across task epochs, and we calculated, for each participant, the min–max spread along each axis for each of our four task epochs (Pre-, Early-, Late-, and Post-learning) and performed an rmANOVA on these values. This analysis revealed a significant effect of epoch along the visual-objects axis (F(3,96) = 5.10, p = 0.008), indicating changes in the distribution of regional embeddings along this axis during SL (see Hong et al., 2019 for a similar approach). Notably, when we performed this same analysis for the autobiographical-episodic (PC2) and working-task (PC3) axes, we observed no significant changes in their expression across epochs (autobiographical-episodic axis: F(3,96) = 0.61, p = 0.61; working-task axis: F(3,96) = 2.47, p = 0.10), indicating dynamic, widespread changes in the distribution and reorganization of regional embeddings during SL and later, when SL is interrupted (see Hong et al., 2019 for a similar approach).
Figure 4B highlights pairs of brain regions (left and right hemisphere) from three distinct networks that load positively onto the visual-object axis (visual network, purple; DAN, green; PRC, black). Together, these regions indicate that there appear to be large, yet differentiable, shifts in the embedding of certain regions across epochs. For example, the visual network and DAN exhibit large shifts from Early- to Late-learning (Fig. 4B, middle), whereas the PRC exhibits its largest shifts from Late- to Post-learning (Fig. 4B, right). To test for these and any other regional effects, we performed an rmANOVA for each region's loadings onto the visual-object axis across task epochs. Following FDR correction (Benjamini and Yekutieli, 2005; q < 0.05), we found that 129 regions exhibited a significant main effect of epoch (Fig. 5A). These regions were predominantly within two major cortical networks, the visual network (63 regions) and DAN (37 regions), with smaller clusters also found in the DMN (9 regions), cingulo-opercular network (8 regions), somatomotor network (6 regions), limbic network (2 regions), and Salience and Ventral Attention (SalVentAttn) network (1 region). Notably, we also found significant areas in the MTL, including left and right PRC and right ERC (Fig. 5A). Figure 5B plots the trajectory of change for significant regions along the visual-objects axis.
Visual SL is associated with changes along the visual-object connectivity axis. A, Regions displaying significant changes in embedding across task epochs. B, Temporal trajectories of the significant regions from A. Each region is depicted as a single point in the two-dimensional axis space, with significant regions colored according to their functional network assignments on the right (Yeo et al., 2011). Colored circles denote each significant regions embedding during the Pre-learning epoch, and each trace attached to the circle represents that region's displacement along the two axes throughout the subsequent epochs (Early-, Late-, and Post-learning). Annotated MTL and hippocampal regions are denoted at the bottom, with colors corresponding to their respective regions. C–E, Top, Unthresholded maps of pairwise comparisons between task epochs. Outlined regions denote significant pairwise comparisons. Positive (red) values indicate increases in loadings (expansion along the axis) and negative (blue) values indicate decreases (contraction along the axis). Bottom, Polar plots depicting mean functional network loading differences for pairwise comparisons. Positive (red) and negative (blue) dots reflect average increases and decreases along the axis. The color behind each brain indicates its functional network assignment (legend in B), with letters depicting its subnetwork assignment (Yeo et al., 2011). Asterisks indicate FDR-corrected significance in network changes. For results demonstrating our findings across different thresholding parameters, see Extended Data Figure 5-1.
Figure 5-1
The main experimental effects in our study are impervious to changes in functional connectivity matrix thresholding parameters. (A) Pearson correlations along the visual-objects axis using row-wise thresholding parameters of 20%, 15%, 10% (original), and 5% in our preprocessing pipeline. (B) Pearson correlations between unthresholded pairwise comparisons across task epochs of interest using the same thresholding parameters as A. The dotted gray line in each scatter plot indicates the original 10% thresholding parameter used in the main analysis. Asterisks indicate FDR-corrected significance in Pearson correlations (as compared to zero). Download Figure 5-1, TIF file.
To characterize these main effects, for each significant brain region we performed follow-up FDR-corrected (Benjamini and Yekutieli, 2005; q < 0.05) paired-samples t tests between epochs and also aggregated across regions (regardless of their significance) within each functional network (Yeo et al., 2011). To test for significant changes in loading along the visual-object axis at the onset of learning, we performed an Early-learning > Pre-learning contrast. This contrast revealed contractions within the visual cortex and DAN (outlined in blue; Fig. 5C, top), which could also be observed qualitatively when aggregated at the level of functional network (Fig. 5C, bottom). Note, however, that the network-level changes in loadings between the Pre- and Early-learning epochs were not significant following FDR correction (also note that, for the polar plots, we used the 17-network cortical mapping in addition to mapping the hippocampus and MTL separately to improve spatial precision compared with the 7-network mapping; Yeo et al., 2011; Wisse et al., 2016; Schaefer et al., 2018).
More critically, to examine changes from the onset of learning to the end of SL, we performed a contrast of Late-learning > Early-learning. This contrast revealed that regions within the visual cortex, which had originally contracted, now significantly expanded along the visual-object axis (Fig. 5D, top, red areas), whereas areas of the DAN contracted further (Fig. 5, compare D, top; C, top). However, at the whole-brain functional network level, we found that only the visual network showed significant changes from Early- to Late-learning (Fig. 5D, bottom).
Although these above neural changes from Early- to Late-learning are interpreted as reflecting an intrinsic process related to SL, an alternative explanation could be that they are merely attributable to subject fatigue or habituation (i.e., the passage of time). Behavior from the perceptual cover task, however, suggests this alternative explanation is unlikely, as participants displayed faster reaction times and greater accuracy from both Early- to Late-Learning and from Pre- to Post-learning (both ps < 0.05). Nevertheless, to further rule out this alternative at the neural level, and to determine how the brain changes when SL-derived associations are disrupted, we examined whether the observed pattern of changes (i.e., expansion of visual cortex regions) also persisted into the final, Post-learning epoch. When we contrasted the Post-learning > Late-learning epochs, we instead found that regions within the MTL—specifically the PRC and ERC—as well as the retrosplenial cortex which did not exhibit significant changes during SL, now modified their connectivity and significantly contracted along the visual-object axis (Fig. 5E, top). In addition to these MTL regions, sparse areas of the visual network contracted, and areas of the DAN expanded. In both cases, this is a reversed trajectory, opposite to how these regions changed during learning. When extrapolated to the network level, there was no change in visual or DAN but resulted in significant changes in both the MTL and DMN-C (Fig. 5E, bottom). This finding provides strong support for the interpretation that the neural changes described earlier reflect processes intrinsic to SL rather than mere time-dependent effects. Moreover, it highlights the widespread network reorganization that occurs when learning is interrupted.
Alterations in regional connectivity that underlie changes along the visual-objects axis
While significant changes in the embedding of regions along the visual-objects axis indicate a change to their overall connectivity profile, this does not itself reveal how these regions change their connectivity with other areas. To elucidate these changes, we performed seed connectivity analyses using representative regions from left and right hemispheres of the visual network, DAN, and PRC. Together, these regions capture the main patterns of effects observed with respect to the expansions and contractions along the visual-objects axis (see Fig. 6 for seed analyses of the right hemisphere; see Extended Data Fig. 6-1 for the left hemisphere). Importantly, we used the same seed regions highlighted in Figure 4B, allowing us to directly link changes in whole-brain functional connectivity to displacements along the visual-object axis.
Patterns of connectivity underlying changes along the visual-objects axis throughout visual SL. A, Connectivity changes of representative seed regions in right visual, DAN, and PRC areas (shown in green, as indicated by green arrow) from Late-learning > Early-learning and Post-learning > Late-learning. Positive (red) values show increases and negative (blue) values show decreases in connectivity. B, Polar plots show connectivity changes between epochs at the network level (according to the Yeo 17-networks parcellation; Yeo et al., 2011; Schaefer et al., 2018), as well as MTL regions derived from the Automated Segmentation of Hippocampal Subfields (ASHS; Wisse et al., 2016) split into MTL (M) and hippocampus (H). The color behind each brain indicates its functional network assignment (legend in B), with letters depicting its subnetwork assignment (Yeo et al., 2011). Asterisks indicate FDR-corrected significance in network changes between epochs. For connectivity changes in the left visual, DAN, and PRC seed areas, see Extended Data Figure 6-1.
Figure 6-1
Patterns of connectivity underlying changes along the visual-objects axis throughout visual SL. (A) Connectivity changes of representative seed regions in left visual, DAN, and PRC areas (shown in green, as indicated by green arrow) from Late-learning > Early-learning and Post-learning > Late-learning. Positive (red) values show increases and negative (blue) values show decreases in loading across along the visual-objects axis. (B) Polar plots show seed-based changes in connectivity between epochs at the network level (according to the Yeo 17-networks parcellation (Thomas Yeo et al., 2011; Schaefer et al., 2018), as well as the additional MTL regions derived from the Automated Segmentation of Hippocampal Subfields (ASHS; Wisse et al., 2016) split into MTL (M) and hippocampus (H)). The color behind each brain indicates its functional network assignment (legend in B), with letters depicting its constituent subnetwork assignment (Thomas Yeo et al., 2011). Asterisks indicate FDR-corrected significance in network changes between epochs. Download Figure 6-1, TIF file.
For each seed region, we contrasted the associated whole-brain connectivity maps for Early- to Late- and Late- to Post-learning with region-wise paired-samples t tests. In Figure 6 we display (1) the unthresholded region-wise contrast maps for each seed region to allow for a complete visualization of a given region's connectivity changes (Fig. 6, left) and (2) corresponding polar plots, depicting the network-level changes in connectivity for each region (Fig. 6, right). Together, these plots provide a comprehensive description of the collective changes in connectivity that contribute to the changes in regional embedding along the visual-objects axis.
For the right visual seed region (Fig. 6A), which exhibited significant expansion along the visual-object axis from Early- to Late-learning (Fig. 5D), connectivity during SL was primarily characterized by increased covariance with other visual network areas, as well as with regions in the DMN and limbic network. In contrast, from Late- to Post-learning, this connectivity pattern reversed, with decreased covariance in these initial areas and the MTL, and increased covariance with the SalVentAttn network. Although these reversals in connectivity did not result in a statistically significant contraction of this seed region along the visual-object axis during Post-learning, this general region of cortex did tend to exhibit contraction during this Post-learning period (see negative changes in blue in Fig. 5E). Together, these results suggest that, during SL, the visual network's expansion along the visual-objects axis is mainly driven by increased within-network connectivity; however, once SL is interrupted, these patterns of connectivity appear to reverse, with visual cortex exhibiting increased connectivity with brain networks linked to exogenous attention (SalVentAttn network).
For the right DAN (Fig. 6B), which exhibited significant contraction along the visual-object axis from Early- to Late-learning (Fig. 5D), we found that its connectivity during SL was most characterized by a decrease in connectivity with other DAN areas and the visual cortex and an increase in connectivity with higher-order transmodal regions belonging to the DMN B, Control B, and SalVentAttn B subnetworks. Conversely, from Late- to Post-learning, when this region significantly expanded along the visual-object axis, we observed a stark reversal in connectivity patterns. The DAN region now showed increased connectivity with areas within its own network, as well as with regions of the visual network. These findings suggest that the DAN's contraction during learning is driven by decreased within-network connectivity and increased between-network connectivity with higher-order brain networks. In contrast, its expansion during Post-learning is characterized by increased within-network connectivity and reinstated connectivity with the visual cortex.
Finally, for the right PRC (Fig. 6C), which exhibited significant contraction along the visual-object axis from Late- to Post-learning, our seed connectivity maps revealed that, during Post-learning, the PRC decreased connectivity with the visual cortex, while increasing connectivity with large swaths of parietal and prefrontal cortex. The fact that the PRC only significantly changed its embeddings along the visual-object axis once SL was interrupted suggests that it is responsive to prediction errors and may communicate these errors with the parietal and prefrontal cortex.
Collectively, these seed-based analyses allow us to characterize, at the level of whole-brain functional connectivity, the selective changes in areas along the visual-objects axis both during SL and once the learned associations are violated (during Post-learning). The findings generally support the interpretation that expansions along the visual-object axis reflect, in part, increased within-network connectivity, while contractions indicate increased between-network connectivity.
The role of the hippocampus in statistical learning
In previous sections, we characterized regional changes along the visual-objects axis that occur in the visual network, DAN, and PRC during visual SL, as well as the underlying shifts in functional connectivity driving these changes. While our findings in the PRC suggest its role in processing prediction errors, we observed no significant changes in the embedding of hippocampal regions across our task—a conspicuous absence given prior work implicating hippocampus in SL (Schapiro et al., 2012, 2014) and evidence for representational change in our earlier analyses (Wammes et al., 2022; Fig. 2C). One explanation for this discrepancy may stem from our functional connectivity analysis including 998 cortical but only eight hippocampal regions. Since dimension reduction approaches like PCA aim to explain the largest amount of variance in a relatively small number of orthogonal components, it is not surprising that our strongest loadings (PCs 1–3) were dominated by cortex and not the hippocampus (Fig. 3A). It may also be that the changes that occur in the hippocampus are largely local and intra-hippocampal and not observable using this higher-level analysis that predominantly focuses on macroscale changes in connectivity across networks. Nevertheless, as prior work has identified a clear role for the hippocampus in SL, we sought to directly examine hippocampal connectivity changes during and after learning. To this end, we aggregated seed-based hippocampal connectivity maps for subfields in left and right CA1, CA2/3, DG, and subiculum and created contrast maps to compare Early- to Late-, and Late- to Post-learning.
Our hippocampal-based seed connectivity analysis (Fig. 7A; for the left hippocampus, see Extended Data Fig. 7-1) revealed that from Early- to Late-learning, the right hippocampus showed modest qualitative increases in connectivity with all other networks, but specifically with other areas of the MTL, and limbic-A subnetwork (Fig. 7B). However, once SL was interrupted during Post-learning, these patterns tended to reverse, with the hippocampus most notably decreasing its connectivity with the visual cortex. In addition, the hippocampus also increased connectivity with the contralateral hemisphere and the retrosplenial subdivision of the DMN (i.e., DMN-C subnetwork). While this reversal in connectivity aligns with changes within the PRC and the overall decrease along the visual-objects axis (Fig. 5E), it is notable that no significant changes were observed in the hippocampus. This, in conjunction with the striking patterns seen in other MTL regions, suggests that the PRC and ERC may play a more crucial role than previously thought in mediating information exchange between the cortex and hippocampus during learning and when the learned associations are violated.
Hippocampal-based connectivity analysis. A, Connectivity changes for the right hippocampus (in green) from Late-learning > Early-learning and Post-learning > Late-learning. Positive (red) and negative (blue) values show increases and decreases in connectivity, respectively. B, Polar plots show seed-based changes in connectivity between epochs at the network level (according to the Yeo 17-network parcellation; Yeo et al., 2011; Schaefer et al., 2018), as well as the additional MTL regions derived from the Automated Segmentation of Hippocampal Subfields (ASHS; Wisse et al., 2016) split into MTL (M) and hippocampus (H). The color behind each brain indicates its functional network assignment, with letters depicting its constituent subnetwork assignment (Yeo et al., 2011). For connectivity changes in the left hippocampus, see Extended Data Figure 7-1.
Figure 7-1
Hippocampal-Based Connectivity Analysis. (A) Connectivity changes for the left hippocampus (in green) from Late-learning > Early-learning and Post-learning > Late-learning. Positive (red) and negative (blue) values show increases and decreases in connectivity, respectively. (B) Polar plots show seed-based changes in connectivity between epochs at the network level (according to the Yeo 17-networks parcellation; Thomas Yeo et al., 2011; Schaefer et al., 2018), as well as the additional MTL regions derived from the Automated Segmentation of Hippocampal Subfields (ASHS; Wisse et al., 2016) split into MTL (M) and hippocampus (H)). The color behind each brain indicates its functional network assignment, with letters depicting its constituent subnetwork assignment (Thomas Yeo et al., 2011). Download Figure 7-1, TIF file.
Discussion
The current study aimed to characterize changes in the landscape of whole-brain functional activity that underlie visual SL. We used recent advances in dimension reduction techniques to build a low-dimensional manifold which characterized the main axes of whole-brain connectivity. Among these main axes, we found only a single connectivity dimension—the visual-object axis—along which regions significantly altered their expression during learning and subsequent violations of the learned associations. Focused analyses on this axis revealed that during SL, regions within the higher-order visual cortex expanded along this axis, reflecting both their decreased functional connectivity with other networks, and increased within-network covariance. Meanwhile, areas of the DAN contracted along this axis, reflecting their increased functional connectivity with other networks and decreased covariance with the visual cortex. Importantly, when learned associations were violated in Post-learning, we observed a markedly different pattern: MTL regions, previously unaffected during initial learning, exhibited contraction, while the trends in visual and DAN regions either halted or reversed. Together, our findings provide a unique window into the whole-brain changes that underlie visual SL, as well as the violation of learned associations. Our results suggest that, at the macroscale, SL is predominantly characterized by cortical connectivity changes, whereas the MTL, particularly extrahippocampal regions like the PRC and ERC, respond to violations in the learned associations. This suggests a nuanced interplay between cortical and subcortical structures in supporting SL, providing a different perspective on the neural underpinnings of this fundamental cognitive process.
From Early- to Late-learning, we found that regions within the visual network significantly expanded along the visual-objects axis. This expansion was primarily characterized by increased within-network connectivity (i.e., modularity), likely reflecting a mechanism whereby the modularity of task-specific networks is increased to insulate encoded associations from additional change (Kashtan and Alon, 2005; Félix and Wagner, 2008) and thus enhance the robustness of learning. In contrast, the DAN's contraction along the visual-objects axis during SL was driven by decreased connectivity with other DAN regions and the visual cortex and increased connectivity with areas of transmodal cortex (Control and SalVentAttn networks). These findings are consistent with prior work showing reduced connectivity between visual and higher-order association regions as SL takes place (Karuza et al., 2017), suggesting that as regularities are encoded, cross-talk between sensory areas involved in task demands and areas involved in attentional control become less important for learning.
A novel aspect of our task was the inclusion of SL blocks book-ended by unstructured runs, allowing us to uniquely investigate changes in whole-brain activity when learned visual associations were disrupted (in Post-learning). Consistent with the idea that the network connectivity effects were related to visual SL rather than merely the passage of time, we observed large reversals in several of the whole-brain seed connectivity patterns. Indeed, regions that had increased or decreased in covariance with the visual network and DAN during learning now showed opposite patterns after SL was interrupted. Specifically, the visual cortex, which had become more modular during learning, exhibited reduced within-network connectivity and increased connectivity with higher-order attention networks. Similarly, while the DAN became less modular during learning, it now showed increased within-network connectivity and greater connectivity with the visual network. These marked reversals in connectivity suggest that the brain's reorganization during SL is partially undone when learned patterns are disrupted, possibly reflecting a response to a prediction error signal (see below).
One particularly noteworthy finding was that the hippocampus, despite showing theory-consistent representational change as a result of SL, did not exhibit significant changes in its connectivity during or after learning. However, extrahippocampal MTL regions that did not initially show learning effects—the PRC and ERC—now exhibited significant contractions along the visual-objects axis in Post-learning. These contractions reflected an increased connectivity with areas of the DAN, DMN, and higher-order control network, as well as decreased covariance with the visual cortex. Current frameworks highlight the PRC as a specialized region positioned at the apex of the ventral visual stream, supporting object recognition (Buckley, 2005; O’Neil et al., 2009; Schapiro et al., 2012), and the ERC as a bridge between the ventral visual pathway and the hippocampus, via (in part) PRC (Eichenbaum et al., 2007). The PRC and ERC's contraction likely reflects a response to prediction error, whereby these regions recouple with higher-order cortical networks to reorient attention to incoming sensory information. Prediction error signals evoked in the hippocampus (Davachi and DuBrow, 2015; Long et al., 2016) propagate out to cortical regions implicated in perception and attention to update existing schemas (Wang and Morris, 2010; Hindy et al., 2016). When predictions are violated, the outcome is often renewed communication between sensory and association areas (Bein et al., 2021), including between the hippocampus and its sensory inputs (Lisman and Grace, 2005). Our findings not only extend this theoretical view of SL but also provide a novel whole-brain perspective of how such prediction errors may be distributed across the cortex. While prior work has focused on the role of the PRC and ERC in more explicit forms of learning (McClelland et al., 1995; Dickerson and Eichenbaum, 2010), our results suggest a crucial role for these regions in detecting violations in implicit learning, perhaps precipitating the reversals observed in our visual and DAN seed regions. Similarly, while prior work has largely implicated the hippocampus in responding to prediction error, our findings indicate that other MTL regions likely play a larger role in controlling the flow of information than previously assumed.
The differential role of cortical and MTL networks in our main analysis (Fig. 5) provides a more nuanced understanding of whether SL operates primarily through cortical or MTL mechanisms. Specifically, our results highlight that interactions within and between these regions dynamically update during SL, leading to different network configurations depending on the phase of the task. Both cortical and MTL circuits appear to learn in parallel: the hippocampus undergoes local representational change to encode paired stimuli (Schapiro et al., 2012, 2016; Wammes et al., 2022), while sensory cortical circuits become more modular, disconnecting from higher-order cognitive networks. However, when learned regularities are disrupted, the MTL detects this mismatch and increases communication with these cognitive networks, while decreasing its connectivity with the primary visual cortex. Importantly, our current findings suggest that these above changes in the landscape of whole-brain activity are largely independent from the hippocampus, suggesting a more prominent role for other MTL regions, like PRC and EC. This highlights the complexity of SL mechanisms, where cortical and MTL regions may contribute in distinct yet complementary ways. An alternative explanation that is worth considering is that the level at which the hippocampus encodes learned associations may not be captured by our current methods, which focus on larger-scale neural network changes.
Although our manifold-based analyses were necessarily exploratory, they were informed by previous research on cortical and MTL involvement in SL (Huettel et al., 2002; Schapiro et al., 2012; Karuza et al., 2017). By revealing global connectivity changes that align with prior demonstrations of visual and attention networks—and potentially extend to extrahippocampal MTL regions—we generate hypotheses that future, more targeted experiments can directly test. For instance, manipulating the timing or salience of disrupted pairings may reveal whether the PRC's transient re-engagement indeed reflects a prediction error mechanism. Ultimately, we view these data-driven findings as offering a new framework for hypothesizing how entire brain networks reorganize to facilitate, and then update, visual statistical learning.
While the current study offers several unique insights into the whole-brain changes in functional connectivity associated with visual SL, there remain several important targets for future research. First, while our findings point to distinct patterns of connectivity change associated with learning and the later violation of that learning, future work could directly track the behavioral correlates of these neural changes on a more moment-by-moment basis. Second, while the current study explores implicit learning, it remains underexplored how the neural changes observed here may relate to other forms of learning (e.g., auditory, spatial SL, error-based learning), as well as more general distinctions made between implicit versus explicit forms of learning in the literature (Kahneman, 2011). For example, in the sensorimotor control literature, implicit learning is thought to be primarily cerebellar dependent, and yet it is unknown whether statistical learning also relies on cerebellum, which contains similar analogous functional networks to cortex (Choi et al., 2012). Future work will be needed to test these ideas. Third, given its established role in SL and prediction (Shohamy and Wagner, 2008; Van Kesteren et al., 2012), it was surprising to us that the hippocampus did not significantly change its embedding along the visual-objects axis during any phases of the task. Nevertheless, when we conducted a seed connectivity analysis focused on the hippocampus directly, we observed a muted pattern of effects, comparable with what was observed in the PRC. In future work, it may be beneficial to construct a hippocampus-centric manifold to increase the sensitivity in detecting changes in its patterns of covariance; alternatively, our findings and whole-brain approach may highlight the relatively greater importance of other regions (outside the hippocampus) in guiding SL and the encoding of violations of learned associations. Finally, future work could incorporate other analytic methods, such as graph theory, to provide complementary perspectives onto the large-scale, functional changes we observed across visual SL using PCA. Alternative approaches such as these would broaden our understanding of how whole-brain cortical and subcortical interactions evolve over the course of learning, and later, once learning is disrupted.
In summary, our results provide a unique whole-brain perspective for understanding connectivity changes during SL and suggest that while cortical and MTL regions are involved in SL, this may occur via different mechanisms and during different time periods. That is, cortical changes subserve learning through repetition, while MTL handles reorganization to allow for updating in response to unexpected deviations from prior learning.
Footnotes
This work was supported by the Canadian Institutes of Health Research PJT175012 and Natural Sciences and Engineering Research Council of Canada RGPIN-2017-04684.
↵*J.P.G. and J.D.W. contributed equally to this work.
The authors declare no competing financial interests. D.J.G. and J.P.G. are employees of Voxel AI, but this research was not funded by Voxel AI, nor does the company stand to gain from the publication of this work.
- Correspondence should be addressed to Jeffrey Wammes at jeffrey.wammes{at}queensu.ca.
