Neural correlates of learning speed reveal developmental differences in memory Authors: Tess Allegra Forest, Margaret L. Schlichting, Amy S. Finn Published: 2026 ABSTRACT A great deal of research has tried to understand the nature of children’s memory representations across time—especially in regard to how young learners navigate the conflicting pressures of remembering specific details of their experiences versus learning about more general aspects of their worlds. Here, we conducted a preregistered, exploratory analysis of developmental fMRI data to ask how learning speed impacts memory formation, and probe the idea that developmental differences in the balance of specific and general memory might be partially explained by the rate at which children (N = 35, ages 9–10) and adults (N = 35) learn. Using a statistical learning task and a novel application of decoding with MVPA, we quantified whether each learner’s brain detected the presence of structured information early or late in learning. Adults who showed early neural sensitivity to structure formed only general memories while those who detected structure later formed both specific and general memories. In children, however, we observed a trade-off: children who showed early neural sensitivity to structure formed only specific memories while those who detected structure later formed only general memories. These data have important implications for theories of learning and memory development in many contexts. Children remember remarkable, but often peculiar, things about their experiences. Why is it that young children frequently can’t tell you what they did before lunch, but are able to recount highly specific details about seemingly random events, like the exact spot a snail crossed the sidewalk months ago? Over the years, scientists have highlighted that beyond memory simply getting better with age, the types of things children and adults remember are also quite different (Brainerd et al., 2008, Finn et al., 2016, Gualtieri and Finn, 2022, Raj and Bell, 2010, Sloutsky and Fisher, 2004, Wilburn and Feeney, 2008). Even basic learning processes, like the ability to learn about the structure of our environments (termed “statistical learning”), result in different memories with age (Forest, Schlichting, et al., 2023), especially in regard to how children manage the tension between recalling precise details versus general patterns in their experiences. While it is clear that many developmental changes contribute to these differences (Finn and Hudson Kam, 2008, Forest and Amso, 2023, Gómez, 2016, Newport, 1990, Schlichting et al., 2017), we propose that one important developmental change contributing to memory differences may be the rate of learning. To test this, we used a novel, exploratory measure of participants’ rate of learning which measured when during learning an individual person’s brain detected structure during a statistical learning task. Then, we asked if individual differences in the timing of structure detection were related to developmental differences in the emergence of specific and general memories, with broad implications for understanding the development of learning and memory. Decades of past research has made clear that humans—of all ages—learn about structure in the world to support later behaviors via a mechanism called “statistical learning” (Aslin, 2017, Saffran et al., 1996). Statistical learning is supported by a host of brain regions (Karuza et al., 2013, McNealy et al., 2006, Schapiro et al., 2014, Turk-Browne et al., 2009) and neurocomputational processes (Conway, 2020, Schapiro et al., 2017), which (at least in adults) results in a variety of memories being formed on different timescales. For instance, learners form both implicit and explicit memories for structure (Batterink et al., 2015, Bays et al., 2016, Kim et al., 2009, Liu et al., 2023, Siegelman et al., 2018), which decay at different rates (Liu et al., 2023). Even the type of information that is maintained in memory following statistical learning is multifaceted (Endress and Langus, 2017, Endress and Mehler, 2009, Forest et al., 2022, Isbilen et al., 2017, Perruchet, 2018, Perruchet and Pacton, 2006, Slone and Johnson, 2018, Tummeltshammer et al., 2017). Particularly relevant here, adults store both specific information about the exact transitions they witness, and more general, high-level knowledge about the aspects of their experience that tend to happen together during learning (Endress and Langus, 2017, Forest et al., 2022, Isbilen et al., 2017, Koelsch et al., 2016, Schapiro et al., 2013, Schapiro et al., 2016, Sherman et al., 2023). But, these different representations emerge at different points in the learning process (also see Henin et al., 2021)—with both specific memory and high-level general memory not being represented in tandem until after about two minutes of exposure (Forest et al., 2022), which is a relatively long time in the realm of statistical learning tasks. This work all hints that the time spent learning shapes the nature of the resulting statistical memories, in line with research from other fields (e.g. Ahmad et al., 2017). There is reason to think the memory representations formed during statistical learning change a great deal with age (Forest et al., 2023, Gómez, 2016, Jung et al., 2020, Raviv and Arnon, 2018, Schlichting et al., 2017, Shufaniya and Arnon, 2018). For instance, it is not until around age eight that children show both the specific and higher-level flexible memories that adults do (Forest, Abolghasem, et al., 2023). Earlier in life, infants seem to mostly remember general information about items that went together, independent of their order (Gómez, 2016, Slone and Johnson, 2018, Tummeltshammer et al., 2017), while young children look very different, only storing specific information for exact transitions (Forest, Abolghasem, et al., 2023). Piecing these different works together, we observe a developmental trajectory where early in life, general representations of statistical structure guide behavior, followed by specific representations guiding behavior through mid-childhood, followed by both general and specific representations simultaneously guiding behavior in adulthood (Forest, Schlichting, et al., 2023). Across development, we therefore appear to go through shifts in the type of memories we form for statistical structure before a mature pattern emerges. If specific and general memories emerge at different times during learning in adults, one possible explanation for these developmental shifts might be that younger people learn at a different rate; in other words, the amount of experience required to become sensitive to structure might differ. Although prior work has suggested that statistical learning occurs at a similar pace across development (Choi et al., 2020, Moreau et al., 2022, Smalle and Bogaerts, 2024), these studies have tended to average across participants in a group. And, while these prior works have largely measured when during learning participants’ neural activity or reaction times indicate having learned a structure, they have not also measured differences in the type of memories that participants form as a function of time. One possibility for reconciling the lack of developmental differences in the rate of learning with development differences in memory specificity is to look more closely at individual variability in the rate of learning across development. Before we can make clear predictions about how the rate of learning might impact specific and general memory, it is important to note that, outside the context of statistical learning, a great deal of research has tried to understand the nature of children’s memory representations across development—especially in regard to how young learners navigate the conflicting pressures of remembering specific details of their experiences versus learning about more general aspects of their worlds (see Sloutsky et al., 2024 for a recent review). Although there is active debate about whether children’s memories become more specific or more general with age, an important take-away from this literature is that there seem to be (at least) two routes to forming general memories: one that comes at the cost of specific details and another which does not. First, early in life, learners seem to show fuzzy, general memories for the commonalities across related experiences (Brainerd and Reyna, 2004, Brainerd and Reyna, 2013, Gómez and Edgin, 2016), which are formed over many repetitions (Gómez, 2016, McClelland et al., 1995). These memories are likely supported by early-developing cortical learning mechanisms (Gómez, 2016, McClelland et al., 1995), or the CA1 subregion of the hippocampus (Ellis et al., 2021, Schapiro et al., 2017). Then, with further maturation, children come to remember increasingly specific details of unique experiences across childhood (e.g. Canada et al., 2019; Keresztes et al., 2017, Keresztes et al., 2018), supporting the perspective that memory moves from general to specific with age. However, a second important transformation has been suggested to occur later in development—when children begin to be able to simultaneously store the details of unique experiences while also abstracting or integrating across them to form generalized, flexible representations that allow for efficient action in novel situations (e.g., Bauer et al., 2021; Ngo et al., 2021; Schlichting et al., 2022; Varga et al., 2025). While the exact age this may start to appear depends on the study and task, this process is thought to be supported, at least in part, by the relatively protracted maturation (Calabro et al., 2020) of hippocampal-prefrontal connections (Varga et al., 2025), and seems to be in place at least by young adulthood (Bowman et al., 2020, Frank et al., 2019, Schlichting et al., 2015), supporting the idea that memories also become more general with age, from sometime in adolescence or late childhood. The two hypothesized routes to forming general memories, which emerge at different times across development (one very early and one quite late), have clear implications for shaping general and specific memory behavior differently: there is a developmentally early but slow to construct (requiring multiple exposures) route to generalization, which involves learning general information at the cost of specific details and is present from early in life; and there is a developmentally late (around adolescence), but fast to construct (requiring fewer exposures) and more integrative route to generalization which allows learners to simultaneously remember specific details and higher-order structure. Critically, both routes of generalization are available to adolescents and up, meaning that they do not have to make a tradeoff when forming general and specific memories while children may (as discussed in more detail in Forest, Schlichting, et al., 2023). Observing these developmental differences in general memory formation has been difficult, because it requires understanding the rate at which different aged humans learn from similar experiences while also probing the nature of the memory representations that each person forms (general and/or specific). Here, we had a unique opportunity to conduct an exploratory analysis of developmental functional MRI data collected during statistical learning—a task especially well-suited to probing this question because it can result in both specific and general memory, requires repeated experience which allows for measuring when someone learns, and is equally feasible for learners of all ages. This allowed us to ask 1) whether we could measure when children and adults show neural evidence of detecting structure during learning using an MVPA decoding approach as an indirect neural measure of the rate of learning and 2) whether the timing of structure detection relates to memory for both specific and general information across ages. To anticipate, we achieved both goals. Given the reviewed evidence that younger learners might rely on a slower route to generalization but that children often show highly specific memory, we hypothesized that for younger learners, fast structure detection would result in specific memory, while slow structure detection should produce general memory at the cost of specific details. Older learners, who instead might use a more integrative route to generalization, will be able to maintain both simultaneously, although we did not have strong a priori predictions about how this would relate to the timing of learning. This work has important implications both for theories of memory development and modern descriptions of how statistical learning works across the lifespan. Methods Participants 37 young adults (18–35 years, Mage = 23.35 years, 11 Male) and 51 children (9–10 years, Mage = 9.93 years, 24 male) participated in this study at the Toronto Neuroimaging Center. Children visited twice—once for a mock scan and, contingent on mock scan performance, for a scan between one and seven days later. Of those originally recruited, one adult and 15 children did not complete MRI data collection, leaving 36 adults and 36 children for analysis. After applying our pre-registered motion-based exclusion criteria (https://osf.io/fbyr8, see below), 35 adults (Mage = 23.43, 10 Male) and 35 children (Mage = 9.97 years, 17 male) provided data for the analyses presented here. All adult participants provided consent, and all child participants’ parents provided consent. All children also provided assent. Design and stimuli Participants completed a statistical learning phase, a test of their statistical learning, and a subsequent blocked re-exposure to structured versus random information (hereafter, ‘structure versus random phase’) while in the scanner (Fig. 1). Both prior to and immediately after the learning and test phases, participants also watched the same objects (Hsu et al., 2014, Schlichting et al., 2015) used in the statistical learning task appear one-at-a-time in a random order while in the scanner. These data were used to address a separate question, and are not included in this paper.Fig. 1Task and Analysis Schematic. In the scanner, participants completed a structural scan before a variety of other tasks, including a) one run of a statistical learning phase which included four consistently ordered triplets (three visualized) which appeared 24 times each in a pseudorandom order, b) a test phase (not scanned) which included four unique types of test trials (note that while there were multiple possible arrangements of items in Shuffle and Control test trials, just one is visualized for each) and c) a structure versus random phase that included 4 runs of a blocked task in which they watched blocks of images in the same structured order they had seen previously, the same images in a random order, or a fixation cross. The right-hand panel depicts the analysis approach for the data from each phase (grey dashed borders). For behavioral data collected during the test phase (b), Old-New memory responses for different test types were subtracted to index specific (old-shuffled) and general (shuffled-control) memories. For our neural analyses, a linear Support Vector Machine classifier was trained on the blocked data from panel (c) using a leave-one-run-out cross-validation within-participant, and then applied to each timepoint (every two-second repetition time (TR)) of the learning phase data in (a). Task and Analysis Schematic. In the scanner, participants completed a structural scan before a variety of other tasks, including a) one run of a statistical learning phase which included four consistently ordered triplets (three visualized) which appeared 24 times each in a pseudorandom order, b) a test phase (not scanned) which included four unique types of test trials (note that while there were multiple possible arrangements of items in Shuffle and Control test trials, just one is visualized for each) and c) a structure versus random phase that included 4 runs of a blocked task in which they watched blocks of images in the same structured order they had seen previously, the same images in a random order, or a fixation cross. The right-hand panel depicts the analysis approach for the data from each phase (grey dashed borders). For behavioral data collected during the test phase (b), Old-New memory responses for different test types were subtracted to index specific (old-shuffled) and general (shuffled-control) memories. For our neural analyses, a linear Support Vector Machine classifier was trained on the blocked data from panel (c) using a leave-one-run-out cross-validation within-participant, and then applied to each timepoint (every two-second repetition time (TR)) of the learning phase data in (a). Learning phase During the statistical learning phase, participants watched a stream of 12 novel objects (Hsu et al., 2014, Schlichting et al., 2015) which appeared one at a time, for 800 ms each (ISI 200 ms). Unbeknownst to participants, these 12 objects were grouped into four unique and consistently-ordered triplets (e.g. ABC, DEF, GHI, JKL, Fig. 1a). Triplets appeared in a pseudorandom order 25 times each during the one-block learning phase (approximately eight minutes) with the caveat that no triplet could immediately follow itself. Thus, over the course of the learning phase, objects in the same triplet appeared together consistently, while those which spanned triplet boundaries did not. As an example, every time participants saw Triplet 1, they saw Object A, followed by Object B and then Object C, which could then be followed by the first object from Triplet 2, 3, or 4 (D, G, or J). This means the transitional probability (P(X|Y)) between shapes within a triplet was 1.0 while the transitional probability between shapes spanning a triplet boundary was 0.33. This learning phase was identical for every participant. Initial piloting of this task revealed that including a cover task ablated any subsequent memory signal, and so participants were instructed to simply watch the shapes. Test phase Immediately following the learning phase, participants completed a self-paced Old-New memory test while lying in the scanner. No functional MRI data were collected during this phase, which typically lasted about two minutes. On each test trial, participants saw the three objects, one at a time, which were presented using the same timing as the learning phase. They were then asked to endorse that sequence as “Old” or “New”, relative to the pattern they had watched just prior. Participants then rated their confidence in their decision (on a scale of 1–4; The analyses presented below include data from all trials, but restricting analyses to ‘High Confidence’ decisions did not change the results, which can be seen in Supplementary Material Section 2.3). The test included four trial types, designed to assess participants’ memory for different aspects of statistical structure (Described below, Fig. 1b), and test trials were presented in a completely random order, unique to each participant. The first type of test trial (termed “Triplets”) consisted of three objects which had appeared together, in the same order, during exposure (e.g. A-B-C). In other words, these were correct triplets from the exposure phase. The second type (termed “Position matched foils”) were made up of three objects which had been part of different triplets, but which each maintained their within-triplet position relative to exposure (e.g. A-E-I, where A is the first position object from Triplet 1, E is the second position object from Triplet 2, and I is the third position object from Triplet 3). This test type is thus a group of objects participants never saw together during exposure, and which is commonly used in visual statistical learning tasks (Forest et al., 2022, Park et al., 2018, Turk-Browne et al., 2005) against which to compare participants’ responses to Triplets to demonstrate evidence that participants learned the structure they were exposed to. The third test type (“Shuffled foils”) presented three objects which were from the same triplet, but which were rearranged to be in a different, never-before-seen order relative to exposure. There were two possible orderings of objects in these test trials, either B-A-C or A-C-B. If participants store information in their memory about which objects generally appeared together during the exposure phase, independent of the order in which different objects occurred, they might be likely to respond to these items as ‘Old.’ The final test type (“Control sequences”) presented three objects from exposure but neither the within-triplet position nor the triplet-membership information was consistent with the exposure phase. Like the Shuffled foils, these test trials also included two possible orderings (either E-A-I or A-I-E). In this way, the control sequences included the minimum amount of consistent information relative to the exposure phase. Participants answered four test questions of each test type. For each test type, the four test questions collectively used all of the objects from exposure (i.e., each triplet was represented once per test type), resulting in 16 unique test trials total. From these data, we then derived separate indices of Specific and General memory. To measure specific memory, we subtracted each participants’ “Old” responses to Shuffled foils from their “Old” responses to Triplets (Triplets-Shuffles), as we reasoned that any difference in the way participants responded to these test sets must stem from increasingly specific information about the exact order that those three objects appeared in. We operationalized General memory as the difference in “Old” responses to Shuffled foils and Control sequences (Shuffles-Control), as we reasoned that any difference in these test sets must be due to participants knowledge of which objects generally went together even when they were not in the correct order relative to exposure. Structured versus random phase In the final phase of the experiment, participants were asked to watch a stream of objects or view a fixation cross (Fig. 1c). Fixation blocks lasted 12 s. Image presentation blocks lasted 18 s, during which each of the 12 objects from the learning phase appeared once, either in a random order (“random blocks”) or in the same pattern the objects had appeared during the learning phase (“structured blocks”). Thus, the only difference between random and structured blocks was the order in which the objects appeared. In both cases, objects were presented with the same timing as the learning phase. In each of four runs, participants saw four structured, four random, and five fixation blocks. The order of these blocks was consistent across subjects. Across all runs, the structured and random blocks followed each other an equal number of times. MR collection and preprocessing MR data acquisition All imaging data were acquired on a 3.0Tesla Siemens Prisma scanner at the Toronto Neuroimaging Centre. All Functional data were collected in 69 oblique axial slices using an echo planar imaging (EPI) sequence, oriented parallel to the AC–PC axis (repetition time (TR) 2000 ms, echo time (TE) 30 ms, flip angle 70°; 2.0 mm isotropic voxels, multiband acceleration factor of 3). We collected two field maps (TR 888 ms, TE 4.92ms/7.38ms, flip angle 60 °; 2 mm isotropic voxels) to correct magnetic field distortions—one before the first functional run and one about half way through the experiment, immediately following the test phase. We also collected one T1-weighted MPRAGE volume (4.7 ×4.7 ×8.0 mm voxels). MR data preprocessing We preprocessed all MR data using fMRIprep (version 20.0.7, Esteban, Blair, et al., 2019; Esteban, Markiewicz, et al., 2019), using the default settings except where noted (fMRIprep requests that users include an exact copy of the methods text which is generated as part of the analysis pipeline—this exact, verbatim, wording is provided in the Supplementary Material, and we have paraphrased it in the main text). We corrected anatomical T1w images for intensity non-uniformity with N4BiasFieldCorrection distributed with ANTs 2.2.0 (Avants et al., 2008), and used these as a reference throughout the workflow. We then skull-stripped the T1w-reference using antsBrainExtraction, using the OASIS30ANTs as a target template. We segmented brain tissue of cerebrospinal fluid, white-matter, and gray-matter on the brain-extracted T1w using FAST (Zhang et al., 2001). We reconstructed brain surfaces with ReconAll (Dale et al., 1999), and then refined the brain mask with fMRIprep’s custom approach to reconciling ANTs-derived and FreeSurfer-derived segmentations of cortical gray-matter. We performed volume-based spatial normalization to the ICMB 152 Nonlinear Asymmetrical template (Version 2009c, Fonov et al., 2009; MNI152NLin2009cSAsym) via nonlinear registration with antsRegistration (ANTs 2.2.0), using brain-extracted versions of both T1w reference and the T1w template. We also generated a probabilistic, whole-brain grey matter mask for our sample. After all subjects’ grey matter masks were moved into standard space, we overlayed all subjects’ grey matter masks and then filtered by voxels in which at least 50% of participants’ individual masks were labeled as grey matter. This mask was the same for adults and children. Functional data preprocessing For each BOLD run per subject, we corrected functional data for susceptibility distortion estimated from the most recently acquired fieldmap using FUGUE (FSL) and other SDCflows tools, co-registered to the T1w reference using bbregister (FreeSurfer, Greve and Fischl, 2009), which was configured with six degrees of freedom. We estimated head-motion parameters with respect to the BOLD reference prior to spatiotemporal filtering using MCFLIRT (FSL 5.0.9, Jenkinson et al., 2002), and slice-time corrected BOLD runs using 3dshift (AFNI, Cox, 1996). We resampled the BOLD time-series into standard space—first we generated a reference volume and its skull-stripped version using a custom methodology of fMRIPrep. We also estimated a number of confounds from the pre-processed BOLD signal: framewise displacement (FD, Power et al., 2014); DVARS (Power et al., 2014); three global signals from within the CSF, WM, and whole-brain masks; and extracted a set of physiological regressors using temporal and anatomical CompCor variants (CompCor, Behzadi et al., 2007). As a departure from the default fMRIprep settings, we annotated frames that exceeded a threshold of 1.5 mm FD or 1.5 standardized DVARS as motion outliers. Functional data were also spatially smoothed during preprocessing in FEAT (FSL 5.0.9) using a FWHM of 5 mm. Those data were then used as input to our MVPA, described below. In line with our pre-registration, we also checked if any functional runs should be excluded entirely on the basis of motion prior to any analyses. To do this, we additionally flagged the (one) timepoint before and the two timepoints after each fMRIprep-defined outlier as timepoints of concern. For all functional MRI runs, we excluded any run during which more than 30% of the timepoints were flagged. This resulted in one child participant having three runs of useable data and one child participant having two runs of useable data from the structure versus random phase. All 35 adult participants and the other 33 child participants had four runs of useable data. Analysis approach Following preprocessing, we used the data described above to answer two primary questions. First, we used a novel MVPA classification approach to understand how quickly children and adults showed neural evidence of having detected the statistical structure (these analyses were pre-registered, as ‘Analysis Three’, https://osf.io/fbyr8). Next, we were interested in understanding whether the timing of structure detection impacts the memories someone ultimately forms—and whether this differs between children and adults (these analyses were not pre-registered). Behavioral data analyses To understand the structure of adults’ and children’s memory representations we compared responses on the different test types using a 2 × 4 ANOVA with Age Group as a between-subjects factor and Test Type as a within-subjects factor, before following up with paired comparisons where relevant. Then, we used our metrics of specific and general memory to compare the extent to which adults and children expressed either form of memory using a 2 × 2 ANOVA and follow-up comparisons. MRI data analyses To explore the relation between when someone showed neural evidence of having detected the statistical structure and the type of memories they formed, we developed the method visualized in Fig. 1a &1c. As a reminder, the memory test took place at the end of the learning phase for all participants, rather than immediately following structure detection in any particular subject. All classification parameters were decided in accordance with our pre-registration (https://osf.io/fbyr8/?view_only=6cfe710aca8d4146a031aef9a46c5dc5, Analysis Three), using a limited subset of participants’ data (two adults and two children). Importantly, any deviations from our pre-registered analysis plan were made without consideration of the memory data, ensuring we did not bias our classification parameters to facilitate any particular pattern of results. Classifier cross-validation, training, and voxel selection We trained a two-way linear support vector machine (lc-SVM) to determine when participants were viewing structured versus random images during the structured versus random phase of the experiment, in two-stages. The same classification approach was used in both stages: we labelled each timepoint (TR) of the structure versus random phase as either being part of a structured block (“structured”), a random block (“random”), or a fixation block after having accounted for hemodynamic lag by shifting data forward by six seconds (three TRs); we removed any linear trends in each participant’s data; removed timepoints that occurred during fixation blocks; and z-scored the data within-run before training a lc-SVM classifier using the default settings from pyMVPA, without feature selection (Hanke et al., 2009, Pedregosa et al., 2011). In the first stage, we determined which voxels in each participant’s brain were most diagnostic for indicating whether they were viewing structured or random image sequences via a searchlight analysis. To do this, we centered a 4 mm radius sphere on each voxel of the probabilistic grey-matter mask generated for each subject and ran a leave-one-run-out cross-validation analysis in each sphere. We averaged the classification accuracy for the left-out-run across folds, and stored the average accuracy as the value for the center voxel of each sphere. This generated a participant-specific map that reflected which voxels had the highest accuracy in classifying structure and random image presentation. We filtered each participant’s map to only include voxels that were in the top 1% of classification performance, and used them to define a subject-specific best-voxel mask (∼1000 voxels per participant; see the Supplementary Material for visualizations of these best-voxel masks by age group). In the second stage, we trained a new classifier with the identical parameters for each person in their own best-voxel mask, and verified that we could decode structure from random image presentation at an above-chance level in each person in this final classifier. Classifier generalization We next tested the final classifier on each timepoint of the learning phase (which, as a reminder, participants experienced before the structured versus random block phase on which the classifier was trained), in each person’s best-voxel mask. This resulted in a two-second-by-two-second (TR-by-TR) classification of the learning phase, which reflected how much evidence the classifier found in favor of “structure”, ranging from 0% (no evidence for structure/complete evidence for random) to 100% (complete evidence for structure/no evidence for random). In a binary classification, timepoints during which the classifier found more than 50% evidence for “structure” would be classified as “structured” and timepoints during which the classifier found less than 50% evidence for “structure” would be classified as “random”. Critically, all of the images presented to participants during the learning phase were actually presented in a structured, not random, order. Thus, we were interested in identifying when during learning the SVM began to classify the input as structured, with the logic that as the pattern of neural activity became more similar to what it looked like at the end of the experiment once participants had learned the structure, the classifier would find more evidence in favor of structure. In other words, we were interested in when the classifier detected that the participant was viewing structured information. Identifying high-confidence classification Having extracted classifier estimates for how likely a participant was to have been looking at structured images at each timepoint during the learning phase, we next labelled each classification as high or low confidence, labelling the 25% of trials on which the classifier was most confident in its guess (for either structure or random) for each participant as “high confidence”, and all others as “low confidence”. We then restricted our analyses to high confidence trials, therefore excluding trials on which the classifier found only slightly more evidence for structure than random. (Note that the decisions to limit analyses to high confidence trials, and the cut-off used, were not pre-registered, but were chosen prior to subsequent analyses. We report data from two other cutoffs, 20% and 30% in the Supplementary Material, Section 2.1). Consistent with the idea that the pattern of neural activity during learning became more similar to the pattern of neural activity during structure and random viewing over time, the classifier became more confident over the course of the learning phase, as evidenced by a significant effect of trial number on whether a guess was high or low confidence in a binomial regression predicting confidence judgments across time in children and adults (χ12= 22.51, p < 0.001). Note that this does not indicate confidence in structure classification (which is correct for the exposure phase), but indicates confidence for both correct (structure) classification and incorrect (random) classification. However, there were no effects of age group on overall confidence rates (χ12= 0.001, p = 0.99) and no interaction between age group and trial number (χ12= 0.04, p = 0.84). After identifying high confidence classifications, we averaged the classifier’s guesses for each minute of the eight-minute-learning phase (30 TRs). Finally, we categorized participants as early detectors, if their average of confident classifications was more than 51% ‘structured’ in the first minute. If this average did not exceed 51% until the second minute or later, they were labelled late detectors. All participants eventually exceeded this 51% mark (and thus everyone ‘learned’ eventually). This cut-off was informed by multiple prior studies on the time-course of statistical learning which have suggested that learning in adults happens within the first presentation of a pattern (Batterink, 2017), that adults presented with less than about 1.5 min of a pattern only form higher order, general, memories for that structure (Forest et al., 2022), and that infants’ and adults’ brains demonstrate neural entrainment to structure in about 1.5 min (Choi et al., 2020). While these studies all indicate learning happens within the first minute or two, we used a one-minute cut-off as this roughly divided our participants into an even split of early and late detectors. Information about how many additional participants showed evidence of learning in subsequent minutes can be found in the supplementary material (Section 2.2). Finally, to relate the rate at which someone detected structure to their memory, we ran a 2 × 2 between-subjects ANOVA with Age Group (child or adult) and Detection Rate (early or late detector) to examine memory performance on both our specific and general memory metrics, and followed up with paired comparisons where relevant. Results Decoding structured from random image presentation Before looking at the learning phase, we asked if we could reliably decode structured image presentation from random image presentation in both children and adults during the structure versus random phase, an interesting question in its own right given that the only difference in stimuli across random and structured blocks was the order of presentation. We observed reliable decoding in both children and adults in the best-performing voxel mask generated for each subject. Performance was also above chance in all participants individually (minimum χ2 = 14.81, all p’s < 0.002). In children, classification accuracy was 67.8% on average (classification accuracy range = 60–77%). In adults, it was 69.3% (range = 60–80%), both of which were significantly better than chance (one-sample t-test relative to 50%, in children: t(34)= 23.48, p < 0.001, Cohen’s D = 3.97, and in adults: t(34)= 23.47, p < 0.001, Cohen’s D = 3.97). Although this above-chance performance is perhaps unsurprising given the preselection of voxels that are sensitive to structure versus random images, there was, importantly, no difference in our ability to successfully classify structured and random images in children and adults (t(67.55) = −1.35, p = 0.18, Cohen’s D = 0.32). Next, we turned to understanding how the pattern of neural activity was classified at each timepoint (TR) of the learning phase. As noted in the methods, these analyses examine classification over time by first looking at the trials on which the classifier was most confident (top 25%), and then averaging these confident classifications for each minute of the learning phase. Importantly, all participants’ average of confident classifications were above 51% for at least one minute of the learning phase (Mean “structure” guess in highest minute = 73.5%, SD = 10.7%), and the mean “structure” guess in the highest minute did not differ by age (t(64.81) = −1.11, p = 0.27, Cohen’s D = 0.27). On average, children’s neural activity showed the most evidence of “structure” 3.41 min into learning, while for adults this was 4.17 min into learning. Thus, the minute during the learning phase during which neural activity most closely matched the structured image presentation was numerically earlier in children than adults, but not significantly so (t(66.65) = −1.60, p = 0.12, Cohen’s D = 0.38). Finally, we classified learners into those who showed neural patterns most consistent with structured image presentation early in learning and those who did not. To do this, we binned participants into groups according to whether the classifier's average “structure” guess was greater than 51% in the first minute of learning. If it was, that participant was labelled an “Early Detector” and if not, that participant was labeled a “Late Detector.” This cut-off was based on prior literature suggesting statistical learning occurs within the first two minutes of learning for most people (Batterink, 2017, Forest et al., 2022), and a data-driven desire to split learners into roughly equal “Early” and “Late Detecting” groups. This resulted in 15 children being labeled Early Detectors and 20 being labeled Late Detectors, while 18 adults were labeled Early Detectors and 17 were labeled Late Detectors (see Fig. 2a).Fig. 2Visualizations of representative child and adult rate of learning data. For each participant individually, we binned timepoint-by-timepoint (TR-by-TR) responses into low (grey) and high confidence (colored) timepoints (small dots) based on whether the response was in the most confident 25% of responses (i.e. the trials on which the classifier was the most confident) for that individual. Then, we averaged the probability of structure in high confidence timepoints for each minute of the learning phase (larger colored dots). Participants who had an average response of more than a 51% chance of “structure” in the first minute of learning (i.e. the first dot in each graph) were labeled as “Early Detectors” (representative child and adult in light blue and light orange shades, respectively). Participants who did not were labeled as “Late Detectors” (dark blue and dark orange, respectively). Please note these figures plot individual participants’ data, using one representative individual from each group. Visualizations of representative child and adult rate of learning data. For each participant individually, we binned timepoint-by-timepoint (TR-by-TR) responses into low (grey) and high confidence (colored) timepoints (small dots) based on whether the response was in the most confident 25% of responses (i.e. the trials on which the classifier was the most confident) for that individual. Then, we averaged the probability of structure in high confidence timepoints for each minute of the learning phase (larger colored dots). Participants who had an average response of more than a 51% chance of “structure” in the first minute of learning (i.e. the first dot in each graph) were labeled as “Early Detectors” (representative child and adult in light blue and light orange shades, respectively). Participants who did not were labeled as “Late Detectors” (dark blue and dark orange, respectively). Please note these figures plot individual participants’ data, using one representative individual from each group. Both age groups learned the statistical structure Both age groups showed behavioral evidence of having learned the structure by treating previously seen triplets as older than control sequences (Children: t(65.94) = 2.28, p = 0.02, Cohen’s D = 0.55); Adults: t(65.33) = 5.92, p < 0.01, Cohen’s D = 1.44), but differed on which test sequences they called old (Age Group x Test Type ANOVA, Main effect of Age Group, F(1,66) = 5.02, p = 0.03, η2G= 0.02; Main effect of Test Type, F(3198) = 20.21, p < 0.01, η2G = 0.18; Interaction of Age Group and Test Type, F(3198) = 4.47, p < 0.01, η2G = 0.05): Adults endorsed shuffled triplets (which were made up of images from the same triplet, just placed in a different order) as older than novel control sequences (t(65.31) = 4.21, p < 0.01, Cohen’s D = 1.02), while children did not (t(64.82) = 0.22, p = 0.83, Cohen’s D = 0.05). Neither group treated the position matched sequences (images from separate triplets, but which maintained the same within-triplet position as during exposure) as older than control sequences (Children: t(65.98) = −1.03, p = 0.31, Cohen’s D = −0.25; Adults: t(65.80) = −0.46, p = 0.65, Cohen’s D = −0.11). Together, this shows that both adults and children learned the structure, with interesting differences in how each age group treats particular lure items. To further probe these age-related differences, we next asked if past work showing that children form specific memories for statistical structure while adults form both specific and general memories was replicated in this study using our predetermined metrics of specific and general memory. A 2 × 2 ANOVA comparing specific and general memory by Age Group showed a main effect of age (F(1,66) = 5.67, p = 0.02, η2G = 0.03), no main effect of test type (F(1,66) = 0.00, p = 1.0, η2G = 0.0) and no interaction (F(1,66) = 2.31, p = 0.13. η2G = 0.023). Specific memory did not differ between children and adults (t(65.75) = 0.15, p = 0.87, Cohen’s D = 0.04), although adults’ specific memory was greater than 0 (t(34)= 2.42, p = 0.02, Cohen’s D = 0.41) while children’s was only marginally so (t(33)= 1.89, p = 0.07, Cohen’s D = 0.32). General memory was greater in adults than children (t(65.96) = 2.65, p = 0.01, Cohen’s D = 0.64, Fig. 3a), in line with past work (Forest, Abolghasem, et al., 2023). In line with this between group difference, adults’ general memory was above 0 (t(34) = 3.81, p = 0.001, Cohen’s D = 0.64), while children’s was not (t(33) = 0.22, p = 0.83, Cohen’s D = 0.04). We therefore observed evidence for general memory representations alongside the specific in adults, while children predominantly showed evidence for specific memory representations as a group.Fig. 3a) Overall specific and general memory by Age Group (Children in blue, Adults in orange). Children and Adults show specific memory, only Adults show general Memory. b) Specific (left) and General (Right) memory in children (shades of blue) and adults (shades of orange), separately for Early (light) and Late (dark) Detectors. Specific memories are formed by Early-Detecting Children and Late-Detecting Adults, while General memories are formed by Late-Detecting Children and all Adults. a) Overall specific and general memory by Age Group (Children in blue, Adults in orange). Children and Adults show specific memory, only Adults show general Memory. b) Specific (left) and General (Right) memory in children (shades of blue) and adults (shades of orange), separately for Early (light) and Late (dark) Detectors. Specific memories are formed by Early-Detecting Children and Late-Detecting Adults, while General memories are formed by Late-Detecting Children and all Adults. Rate of learning predicts different patterns of memory across development Finally, we asked whether the formation of specific or general memories is impacted by the rate at which someone detected structure during learning, and whether this differs between age groups. Examining specific memory first (Fig. 3b, left), we again found no evidence that children or adults were more likely to form specific memories (no main effect of age: F(1,65) = 0.07, p = 0.79, η2 = 0.001), nor that participants who detected structure early versus late were differentially likely to have specific memories by the end of the exposure phase (no main effect of Detection Rate: F(1,65) = 0.002, p = 0.97, η2 < 0.001). There was, however, an interaction between Age and Detection rate (F(1,65) = 6.65, p = 0.01, η2 = 0.09, Fig. 3b) such that in children independently and in adults independently, there was a marginal effect of detection rate (in children: t(30.33) = 1.75, p = 0.09, Cohen’s D = 0.61; in adults: t(32.97) = −1.91, p = 0.07, Cohen’s D = 0.64). That is, Early Detecting Children and Late Detecting Adults showed evidence of having formed specific memories (Children, t(14) = 2.62, p = 0.02, Cohen’s D = 0.68; Adults, t(16) = 3.27, p = 0.005, Cohen’s D = 0.79), while Late Detecting Children and Early Detecting Adults did not (Children, t(18) = 0.29, p = 0.78, Cohen’s D = 0.07; Adults, t(17) = 0.46, p = 0.65, Cohen’s D = 0.11). This pattern of results suggests that the rate at which children and adults learn relates very differently to the types of memories they form. We next examined the formation of general memory (Fig. 3b, right) and found, as expected, that adults exhibited more general memory than children (main effect of age: F(1,65) = 7.10, p = 0.01, η2 = 0.01). There were no overall differences in general memory between early detectors and late detectors (no main effect of Detection Rate, F(1,65) = 1.2, p = 0.27, η2 = 0.02). However, there was an interaction between Age Group and Detection Rate (F(1,65)= 5.80, p = 0.02, η2 = 0.08, Fig. 3c). In this case, Early Detecting Children showed less evidence of forming general memory than Late Detecting Children (t(24.24) = −2.48, p = 0.02, Cohen’s D = 0.89), whereas there was no significant difference in the general memories formed by Early and Late Detecting Adults (t(32.99) = 0.87, p = 0.39, Cohen’s D = 0.29). Accordingly, when each group was examined alone, Early Detecting Children did not form general memory (t(14) = −1.50, p = 0.16, Cohen’s D = 0.39), while all other subgroups did (Late Detecting Children: t(18) = 2.27, p = 0.04, Cohen’s D = 0.52; Early Detecting Adults: t(17) = 3.26, p = 0.005, Cohen’s D = 0.77; Late Detecting Adults: t(16) = 2.05, p = 0.05, Cohen’s D = 0.50). Taken together, in adults we conceptually replicated past work (Forest et al., 2022) showing that early structure detectors only form general memory, but that late detectors formed both specific and general memory. In children, however, there was a starkly different pattern: early detectors formed only specific memories while late detectors formed only general memories. Discussion Here, we used a unique developmental fMRI dataset to conduct an exploratory analysis of how the rate at which people learn impacts memory formation, to probe the idea that developmental differences in the balance of specific and general memory might be partially explained by rate at which children and adults learn. Using a statistical learning task and a novel application of MVPA decoding, we observed how quickly individual learners’ brains “clued in” to the fact that they were watching structured information. We found that the rate at which children’s and adults’ brains detected structure differentially predicted their memories. Adults who showed early neural sensitivity to structure formed only general memories while those who detected structure later formed both specific and general memories. In children, however, there was a clear trade-off: early detectors formed only specific memories while late detectors formed only general memories. We review our findings in turn and discuss their implications below. First, we conceptually replicated past results that children and adults show similar rates of statistical learning. Past research has shown that infants’ (Choi et al., 2020), children’s (Moreau et al., 2022) and adults’ brains entrain to the frequency of statistically meaningful information at a similar rate. Reaction-time based measures of learning have also suggested there are no dramatic differences in the rate at which children and adults become faster to respond to a predictable item in a sequence (Smalle and Bogaerts, 2024), and that the rate of Hebb repetition learning of predictable sequences is similar across age groups (Smalle et al., 2018). These studies indicated that learning happens quite quickly—on the order of a few minutes for neural entrainment metrics (Choi et al., 2020). We likewise observed no difference between age groups in regard to when during learning our classifier found the most evidence that participants were viewing structured information. On average, this was about three and a half minutes into the learning phase for children and four minutes into the learning phase for adults. Thus, although our approach to querying the rate of learning was quite different from past work, it generated promisingly similar conclusions. While we did not observe differences in the timing of when children and adults showed neural evidence of detecting structure, we did observe differences in the type of memories each age group formed. In particular, we replicated past work showing that, on average, children’s memory for statistical information is highly specific, while adults additionally form high-level, generalized memory for their repeated experiences (Forest, Abolghasem, et al., 2023, although note that the children in this study were slightly older than those who expressed just specific memory in past work). These data therefore add to an increasing body of evidence suggesting childhood might reflect a period of time during which memory is uniquely specific (see Sloutsky 2024 for recent theoretical perspective). For instance, children seem to remember distinct exemplars from category learning tasks at the expense of higher-level, more flexible category knowledge (Deng and Sloutsky, 2016, Sloutsky and Fisher, 2004), and younger learners are less likely than older learners to demonstrate false memories for semantically-related experiences (for reviews, Brainerd et al., 2008; Ghetti, 2008). Recent findings show that children push their neural representations for similar, familiar experiences apart so they are especially distinct (Benear et al., 2022), and that neural specificity predicts memory in childhood (Fandakova et al., 2019). Children also have been shown to use differentiated (or, particularly distinct) hippocampal representations of related events to link experiences in their memory, while adults instead use integrated representations housed in the ventromedial prefrontal cortex (Varga et al., 2025). These findings are theoretically consistent with accounts of brain development which suggest that by early childhood, access to hippocampal learning mechanisms allows for increasingly precise encoding of unique details (Callaghan et al., 2021, Canada et al., 2019, Ghetti and Fandakova, 2020, Keresztes et al., 2018, Lavenex and Banta Lavenex, 2013, Riggins et al., 2018), but immature connections between the medial temporal lobe and ventromedial prefrontal cortex may not yet prompt the generalization across similar events seen in adults (Calabro et al., 2020, Paz-Alonso et al., 2008, Schlichting et al., 2022, Varga et al., 2025). Most excitingly, examining individual differences in the timing of learning revealed developmental shifts in how the rate of learning shapes our memories, and provided an important caveat to the idea that children’s memories are quite specific. First, we found that adults who showed early neural evidence of having detected structure only formed general memories, while later-detecting adults formed both specific and general memories. In children, however, there was a trade-off in when one’s brain showed sensitivity to structure and the memories they formed. Early-detecting children formed only specific memories, while late detecting children formed only general memories. We discuss these unique patterns observed in adults and children in turn. In adults, we found that those who showed neural evidence of having detected structure early during learning only showed evidence of general memory, while those who showed later neural evidence of structure detection formed both specific and general memories. On the surface, this pattern of results conceptually replicates prior work which found that changing the amount of exposure adults had to a stream of statistical information led to differences in the types of memories formed: adults who had a limited amount of exposure just formed general memories, while those with additional exposure formed general and specific memories (Forest et al., 2022). That said, in the current work, we measured all participants’ memory after the same amount of exposure, and so early-learning participants experienced more of the structured information after they initially showed evidence of learning. One possibility is that adults who learned early then disengaged with the task, and therefore show a similar pattern of results to the limited exposure group from prior work. An alternative possibility is that adults who learned the structure quickly did in fact learn specific details, but forgot them before the memory test. These (and likely other) possibilities remain to be worked out in future paradigms. Regardless, the later-learning adults performed quite differently, remembering both specific and general aspects of the structure. Thus, the pattern of results in the later learning adults is in line with theoretical (Sherman, Turk-Browne, et al., 2023), computational (McClelland et al., 1995, Schapiro et al., 2017) and empirical arguments (e.g. Bowman et al., 2020; Henin et al., 2021; Schlichting et al., 2015; Sherman, Aljishi, et al., 2023) that adults are able to maintain both types of memory representations in tandem. A different pattern of results emerged in the children: those who detected the structure early remembered specific aspects of that structure (mirroring the group-level results), but those who detected the structure later, remembered only general aspects of that structure. While preliminary, these findings provide support for the proposal that in early life, generalization comes at the expense of specific details, whereas more mature learners can maintain both simultaneously. These findings are in line with other recent work suggesting that generalization behavior is increasingly contingent on specific details with age (Ngo et al., 2021). The data in the present work may also help to resolve an apparent conflict on theories of memory development. One prominent class suggests that children’s memories shift from general to specific with age (Gómez and Edgin, 2016, Keresztes et al., 2018, Vlach, 2014). These theories emphasize that early-life generalization mechanisms require slowly experiencing similar events and learning about the commonalities across them. Data from our slower learners directly support these ideas; these children only showed evidence of having formed general memories. The second class of theories talks about how children’s memories can be uniquely specific (Brainerd et al., 2008, Forest et al., 2023, Plebanek and Sloutsky, 2017, Sloutsky and Fisher, 2004). Data from our fast learners, and the children as a whole group, also support these theories, since they only formed specific memories. The tradeoff we observe here provides a missing piece of data to connect and expand upon these theories and calls for more work looking at developmental shifts in the rate of learning and the consequences this might have for memory representation across development. This also informs longstanding questions about whether children—as compared to older learners—are better or worse at generalization or abstraction. We clearly see that while both adults and some children are forming general memories (as indexed by our behavioral metric), these memories emerge under different circumstances across the two age groups. Younger learners might express general memories at the expense of specifics after more extensive exposure to a structure, while older learners might require less experience and be able to maintain specific details alongside their general knowledge. While this work is theoretically motivated and inspired by long-standing questions about developmental shifts in memory, the techniques we deployed to understand participants’ rates of learning are highly exploratory. We therefore note several important limitations to address in follow-up studies. First, we were primarily interested in successfully training a classifier which was clearly able to differentiate structured from random images in all participants rather than specifying where in the brain that classification should take place. This decision was made prior to pre-registration and in order to have the highest confidence in our classifier’s ability to probe the rate of learning, but it means that in each participant we used a different, small set of voxels. This has a few implications for our results. One, it is important to note that the classifier we trained was optimized for above-chance performance on classifying structured from random images in any given person’s brain, and thus may not generalize well to other subjects or tasks. Two, we are not able to make spatially-specific claims about the rate at which different brain regions show evidence of detecting structure, or which brain regions show age-related change in the time course of learning. This is a top priority for future work in larger datasets. Relatedly, given the relatively limited sample size in this dataset, we opted to use a binary division of participants into Early or Late detectors, rather than a subject-specific measure of when during the learning phase the classifier detected ‘learning’ in a particular person, in order to conserve power via a group-level analysis relating rate of learning to memory, instead of an individual differences approach. That said, individual measures of the timepoint of learning would be interesting to relate to memory metrics in larger samples. Another by-product of this division into Early and Late detectors is that we were not able to carefully characterize other variability in the time course of learning within, or across participants, nor how this related to memory. Future research might attempt to understand what highly variable neural signals means for someone’s later memory representations, or how early evidence of learning followed by a decline might relate to cognitive disengagement or forgetting. Second, these data included children between the ages of eight and ten, and young adults. While we are excited about the implications these data have for theories of memory development across early life, future research would be well-served to carefully measure both learning and forgetting rates in a wider age range to characterize potential shifts in generalization processes more fully, and to better understand how such shifts explain children’s memories (see Knabe et al., 2023, for a recent example), beyond the age band studied here. As an example, one might expect that in ages where detailed memories are less common (toddlerhood, for instance), even relatively quick learners form general memories, if the ability to hang onto specifics is less prevalent. Or, testing learning more immediately after someone learns might allow even toddlers to demonstrate specific knowledge which they otherwise quickly forget. Additional data with a wider age range may also help to clarify whether the early learning adults and slow learning children (who show similar behavioral profiles) may be using similar approaches to generalization. Separately, statistical learning is increasingly recognized to result in both implicit and explicit memories, at least in adults. Given evidence that implicit learning processes develop earlier than explicit (e.g. Bullock Drummey and Newcombe, 1995; Carroll et al., 1985; Finn et al., 2016), related research has also asked whether children’s statistical learning may rely more heavily on implicit processes than adults’. Although results are currently mixed (Janacsek et al., 2012, Moreau et al., 2022, Smalle and Bogaerts, 2024), the extent to which children’s statistical learning is more or less implicit may interact both with their rate of learning and the specificity of their memories in interesting and important ways. This will need to be investigated using additional tasks in future research measuring rate of learning and memory specificity. Future findings aside, the current work presents a novel method of examining the rate of learning across development, which we hope will inspire future research aimed at understanding when different brain regions contribute to learning. The data herein also inform longstanding questions about whether our memories become more specific or more general with age. We join others (Sloutsky et al., 2024) in suggesting that while on the surface, general memory in early life and later in life might be hard to distinguish, the process by which they are formed, and the related fate of specific details, is unique. CRediT authorship contribution statement Margaret L. Schlichting: Writing – review & editing, Supervision, Software, Methodology, Conceptualization. Amy S. Finn: Writing – review & editing, Writing – original draft, Supervision, Resources, Methodology, Funding acquisition, Conceptualization. Tess Allegra Forest: Writing – review & editing, Writing – original draft, Visualization, Software, Methodology, Formal analysis, Data curation, Conceptualization. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.