Analysis of learning curves within mini-blocks of trials demonstrates that participants immediately made use of the abstract task structure in the first segment of the generalization phase. Through the first few trials of each mini-block of the initial and new category training, fMRI participants’ accuracy steadily improved. In contrast, in the generalization phase, accuracy was near ceiling from the first trial (mean p(correct)=0.96, SD = 0.06; Figure 2a). The rate of this learning curve in the generalization phase was significantly lower than either of the other phases (Figure 2b), consistent with participants inferring an adaptive behavior without need for additional experience or feedback (Wilcoxon signed rank tests: Z’s ≥ 3.21, p’s ≤ 0.004, Bonferroni-corrected for multiple comparisons). Similar results were observed in participants who completed the behavioral version of the task in session 2 and passed the generalization accuracy criterion (Figure 2—figure supplement 2a), with a significantly lower rate of change in the accuracy learning curve in the generalization phase compared to the initial training or new category training phases (Wilcoxon signed rank tests: W’s = 45, p’s = 0.01, Bonferroni-corrected for multiple-comparisons).

Figure 2 with 2 supplements see all

Download asset Open asset

Notably, high accuracy during generalization was accompanied by elevated reaction times (RT; mean = 4.04 s, SD = 1.7 s) on the first trial, with a sharp drop-off subsequently (mean = 1.15 s, SD = 0.08 s; Figure 2c). The rate of speeding was significantly steeper in the generalization phase compared to the initial or new category training phases averaged across the first three contexts in each phase (Wilcoxon signed rank tests: Z ≥ 3.46, p≤0.002, Bonferroni corrected for multiple comparisons; Figure 2d). The same pattern in the rate of speeding was observed in participants from the behavioral group who passed the generalization criterion in session 2 (generalization versus initial training RT rates: W = 36, p=0.047, Bonferroni-corrected for multiple-comparisons; Figure 2—figure supplement 2d). Thus, rather than immediately applying a structure learned incidentally during training, participants took additional time to infer values during the first generalization trials – indicative of a concerted, deliberative process.

As each LS was repeated twice via two contexts during generalization, we could test how previously encountering a LS impacted RTs during a repetition. Participants were faster for the second presentation of contexts from the same LS during generalization compared to contexts from different LSs (repeated measures t-tests: (15)≥2.13, p’s < 0.05, d’s ≥ 0.53), indicating that additional inferences were made more quickly when a LS had been accessed previously (Figure 3a). Comparison of the rate of exponential curves fit to these RT data in each participant demonstrated that these curves were steeper in the first context compared to the second context averaged across LSs (Wilcoxon signed rank test: Z = 3.36, p=0.0007), but there were no differences in the rate of these curves according to the sequential order of LSs (Wilcoxon signed rank test: Z ≤ 1.13, p≥0.2, uncorrected; Figure 3b). A numerically similar pattern of results was observed for the rate of change for learning curves in the behavioral group who passed the generalization criterion, but did not reach statistical significance (Figure 3—figure supplement 1). These data argue that participants were faster in completing the transfer of new category values to contexts if a context belonging to the same LS representation had already been activated, but did not necessarily use information about other LSs they had encountered in narrowing this inference problem.

Figure 3 with 1 supplement see all

Download asset Open asset

After an initial inference, participants may have formed rules for responding to the new context-category combinations and no longer used the LS representation throughout the scanned generalization task. We reasoned that if participants are continuing to utilize the latent abstract task structure, this would be evident in switch costs when the active LS changed between-trials (Collins and Frank, 2013). To avoid conflating LS and context switches, we isolated this analysis to only trials where the context changed between trials. We found a significant switch cost in RT for trials when the LS changed versus stayed the same between trials (repeated-measures t-test: t(15) = 2.33, p=0.03, d = 0.58; Figure 2e), which did not differ between scanner sessions (p=0.8). The same effect was independently observed in participants in the behavioral group who passed the generalization criterion (Figure 2—figure supplement 2; t(8) = 3.45, p=0.009, d = 1.15). These data indicate that participants used these LS representations throughout the scanned task. Participants were also faster when both the context and LS remained the same (t(15) = 5.29, p<0.0001, d = 1.32).

Notably, generalization in this task might not only be achieved through the use of a LS representation to bridge contexts. For example, an associative retrieval process could also explain these results if newly trained category values could be transferred through mediated retrieval of contexts held-out of the generalization phase via shared category-value associations learned in the initial training (Kumaran and McClelland, 2012; Schapiro et al., 2017).

To test this possibility, we compared four alternative computational models that generated predictions for participants’ RTs based on different memory network configurations where contexts, category-values, and LSs were represented as nodes with different levels of activation.

First, the conjunctive associative retrieval (CAR) model represented each condition as a combination of category and context features, and linked these conjunctive representations to each other based on shared features, but had no representation of LS (Figure 4a).

Figure 4 with 1 supplement see all

Download asset Open asset

The independent associative retrieval (IAR) model similarly did without a LS representation, but rather than only representing contexts and categories as conjunctions, maintained independent representations of each and linked contexts to each other based on shared category-value associations (Figure 4b).

In contrast to these associative retrieval models, we included two LS models. A simple LS model connected category-values to three LSs, but forewent any representation of context (Figure 4c). A hierarchical latent state (HLS) model connected category-values to contexts, and further clustered these contexts around LSs, so that activation of each context and category was inherited from the LS representation (Figure 4d).

Each of these models relied on the basic assumption that RTs depend on the degree of activation in the network nodes currently activated. On each trial, activation within the network would increase at different rates depending on the relation of nodes to those that were activated during retrieval. Three rate parameters determined this increase for nodes that were directly activated on a trial (α₁), nodes directly connected to those activated (α₂), and the rest of the network which might be searched but was not directly relevant to the current trial (α₃). To establish that these models differed substantially in their predictions, we tested each model on data simulated using the best fit model parameters from itself and the three alternatives. This cross-model comparison clearly demonstrated that the predictions of these four models were dissimilar and each model provided a much better fit to itself than its alternative counterparts (Figure 4—figure supplement 1a).

The HLS model provided the best fit to participants’ RT data in the fMRI group in sessions 1 and 2, and the nine participants in the behavioral group who passed the generalization accuracy criterion in session 2 (Table 1; Figure 4—figure supplement 1a). Comparison of the three rate parameters of this model showed an ordering consistent with the expectation that the degree of activation of items should fall off at a distance from the items presented on the current trial (Figure 4—figure supplement 1b), where values for α₁ were significantly greater than α₂ and α₃ (Wilcoxon signed rank test: Z’s = 3.51, p’s = 0.001, Bonferroni corrected), and values for α₂ were significantly greater than α₃ (Z = 2.68, p=0.02, Bonferroni corrected).

To understand why the HLS model better accounted for participants’ data, we compared RTs simulated from the best fit model parameters for each participant with each of the four models, focusing on data from session 2 for the fMRI group, as it was more pertinent to the neuroimaging results (Figure 4a–d). We organized these data by the order of presentation of each context associated with each LS (as in Figure 3), switches of category mini-block, and all other trials within a mini-block and then compared the root-mean squared deviation (RMSD) of each model for each trial type. We found that both the LS models provided a better fit on category switch trials compared to the CAR model (repeated-measures t-test, t(15) ≥ 4.76, p’s ≤ 0.002, d ≥ 1.19, Bonferroni-corrected for multiple comparisons). Numerically, the HLS model also provided a better fit to these trials than the IAR model, though this difference was not statistically significant (t(15) = 2.02, p=0.06, d = 0.50, uncorrected). The two associative retrieval models tended to systematically overestimate RTs on these category switch trials, likely because category-values associated with the trained context in the new category training phase were only activated by mediated retrieval, whereas in the LS models these category-values benefited from inherited direct retrieval of the higher-order LS representation. The HLS model also had lower RMSD for the first presentation of the second context compared to the LS model (t(15) = 4.76, p=0.001, d = 1.19, Bonferroni-corrected for multiple comparisons). As the LS model had no separate representation of context, any activation related to a prior LS would be fully transferred to its second presentation through another context, resulting in an underestimation of RTs. Instead, the results were more consistent with separate context representations that inherit activation from a higher-order LSs, that is, the use of a HLS structure.

Having verified that participants form and use an abstract task representation to control behavior, we sought to test the neural systems that support this representation. We carried out a representational similarity searchlight analysis (RSA) using multiple linear regression to compare empirical representational dissimilarity matrices (RDMs) from pattern activity to hypothesis RDMs quantifying the predicted distances between conditions based on LSs, contexts, item categories, expected value, interactions between these factors, and control regressors (Figure 5—figure supplement 1).

We found evidence for a LS representation in bilateral dorsal- and ventrolateral prefrontal cortex (PFC) with a rostral distribution, as well as the bilateral precuneus, left middle temporal gyrus, and bilateral inferior parietal lobules (Figure 5a; Supplementary file 1). Comparing this statistical map with resting-state functional networks from Yeo et al., 2011 revealed that this LS representation overlapped most with a frontoparietal network centered on the inferior frontal and intraparietal sulci (Figure 5—figure supplement 2).

Figure 5 with 5 supplements see all

Download asset Open asset

In contrast, context was associated with activity in the bilateral fusiform gyri (Figure 5b). Expected value was associated with activity in OFC, as well as left superior frontal and angular gyri (Figure 5c). Item category was robustly represented throughout visual areas and PFC (caudally on the left and broadly on the right; Figure 5—figure supplement 3).

To test whether these results were consistent across fMRI sessions, we carried out separate whole-brain searchlight RSAs within each session. The resultant statistical maps for the main effects of interest were similar in both sessions (Figure 5—figure supplement 4). Contrasts between statistical maps for all terms in the multiple regression model only revealed a stronger effect for value in the left middle temporal gyrus in session 2 compared to session 3 (MNI coordinates: −48,–22, 0), and no other significant differences in either direction.

As a secondary question, we were also interested in whether higher-order representations of latent task states and expected value influenced each other, or perceptual-semantic representations of item category. To test this, we included regressors for hypothesis RDMs that captured interactions between LSs, value, and item category, as well as control regressors for interactions of these factors with context. Interactions between item category with value and LS were not significant; though items with same value in the same LS were represented more similarly in ventral temporal cortex (VTC) (Figure 5—figure supplement 1), observations recapitulated in ROI-based analyses of activity within the VTC (Figure 6b).

Figure 6 with 1 supplement see all

Download asset Open asset

We also conducted ROI-based RSAs focused on HPC and OFC, given a priori expectation that these regions are involved in representing latent task structure during task performance. RSAs within these ROIs were null for LS representations (Figure 6a,c). To test these effects at a finer grain, we conducted second-level tests of the whole-brain searchlight analysis masked by these ROIs. Here, we found signals consistent with LS and expected value representations in central OFC, but not HPC (Figure 6—figure supplement 1; Supplementary file 2). There were significant effects for category representations in both of these ROIs, in line with broad discrimination between semantically and perceptually distinct categories of stimuli in these regions (Chikazoe et al., 2014; Kuhl et al., 2012; McNamee et al., 2013; Pegors et al., 2015).

Lastly, we tested a univariate contrast of correct and error responses. Although less functionally specific than the above analyses, we expected this contrast to reveal regions broadly involved in engaging with this generalization task. This contrast revealed greater activation in the HPC and OFC on correct responses (Figure 7; Supplementary file 3).

Figure 7

Download asset Open asset

The study design and analyses were pre-registered on the Open Sciences Framework prior to collecting the data (https://osf.io/x6fmb). Participants completed an acquired equivalence task where they learned that nine contexts belonged to three different LSs (three per state) on the basis of their shared value associations for three object categories. They then used this knowledge to complete a generalization task where they inferred values for a new set of categories in contexts with which they had not been previously paired. A schematic overview of the structure of this experiment is provided in Figure 1 and Figure 1—figure supplement 1.

To aid in building a structured representation of the task, participants were told they were playing the role of a photographer selling photos of different item categories in three goblin kingdoms with distinct preferences for these categories. In each trial, participants had to make a bet on whether to sell an image or to pass. This decision was a risky choice where selling could be paid off with a reward in fictitious gold or punished with a loss of gold. After a decision to sell, feedback was provided in terms of the gold won or lost. Passing would lead to no change in gold, but feedback was provided in terms of the outcome had the participant chosen to sell. Thus, in all phases of the experiment except for generalization (see below), fully informative feedback was provided independent of participants’ choices. Participants were incentivized with a monetary bonus proportional to the amount of gold they earned in the task. All photos were trial-unique so that participants never learned the values of specific images, but of general image categories.

This experiment took place over three sessions. Each experimental session was comprised of three main phases. First, in the beginning of the initial training phase, participants were tasked with selling photos of hands, foods, and leaves in three contexts, represented by a natural scene image above the photo they were selling (A1, B1, C1). Participants were told that the contexts each came from a different ‘kingdom’, equivalent to the generative LSs (LS A, LS B, LS C), and that each of these kingdoms preferred one category of items strongly over the other two categories. Within each context one item category was associated with a 90% probability of reward and 10% probability of punishment for a decision to sell, while the other two categories were associated with a 10% chance of reward and 90% chance of punishment. In each trial in this initial training phase, participants could gain 50 gold or lose 25 gold for choosing to sell a photo. Thus, the expected value of selling, without knowledge of the probability of reward for each category, was zero.

After two blocks of this training, participants then encountered ‘roadwork’ after which they would learn about a new context in each kingdom represented by novel natural scenes (A2, B2, C2). This roadwork repeated a second time after another two blocks so that participants learned about three contexts in each of the three kingdoms. Importantly, all contexts associated with a given LS had the same reward probability structure across item categories. Participants were not directly informed of which contexts corresponded to which kingdoms, but had to make these inferences based on their shared category values. The scene images for each context were randomly selected from a set of nine for each participant, so there was no systematic visual relationship between the contexts of each kingdom. Further, as contexts belonging to each ‘kingdom’ were encountered after each episode of roadwork, this event bound was not a cue to task structure and no aspect of temporal adjacency could link contexts to a LS.

These initial training blocks consisted of 54 trials with six trials per condition. This initial training phase was organized in batches consisting of a series of trials for each context-category combination (i.e. a ‘mini-block’). These mini-blocks were nested by context, so that participants saw each category in the same context one after another (e.g. A1-hands, A1-leaves, A1-foods), before switching to a new context and looping once again through each combination of context-category pairs (e.g. B1-leaves, etc.). The order of presentation for these categories within each context was randomized, as was the order of these contexts. Following these six initial training blocks, participants completed a final reminder training block with the three categories presented in all nine contexts in mini-blocks of eight trials (216 trials total).

In the second phase, participants completed training on new image categories (faces, animals, and man-made objects) within a subset of the previous contexts (e.g. A3, B3, C3). Which set of contexts was presented in this phase (i.e. A1, B1, C1 versus A2, B2, C2 or A3, B3, C3) was randomized across participants. Participants completed 90 trials in the course of a single block, consisting of 10 trial mini-blocks for each combination of the three contexts and three categories.

In the new category training phase, two categories of items were now associated with a 90% chance of reward and 10% chance of punishment in each context, while these values were reversed in one context. The categories that were rewarded and punished in each context were randomized with respect to the categories in the initial training across participants. For example, there was no rule in the experiment that faces have high value in contexts where hands have low value. The punishment associated with selling a photo was increased to 100 gold in this block so that the expected value of selling a photo without knowing the probability of reward in a specific category-context combination was still zero.

Finally, in the third phase, participants carried out a generalization test where they were presented with photos from these new categories in conjunction with the six contexts held out of the prior new category training (e.g. A1, B1, C1 and A2, B2, C2, if A3, B3, C3 were trained with the new categories). Critically, participants did not receive feedback throughout this phase and could thus not learn the values of these new categories from experience. Instead, they would have to rely on their knowledge about which contexts belonged to which kingdoms (i.e. LSs) based on the shared category value associations learned in the initial training phase. This generalization test included 10 trial mini-blocks for each of the 18 conditions (each context-category combination, with three categories and six contexts), resulting in 180 trials total. Participants were informed that the value in gold earned and lost on each trial had been doubled from the new category training phase to further incentivize performance.

It is important to emphasize that generalization during this phase was only possible based on a representation of the shared LSs among contexts. The particular pairings of items and contexts during generalization had never been encountered previously. And, as no feedback was provided during this phase, mappings could not be learned through reinforcement. Further, as the item categories themselves were different from the initial training phase, nothing about the objects or values could link the contexts encountered during generalization to what was learned during phases 1 or 2 except for an abstract LS. That LS already clustered the contexts encountered during generalization with that encountered during new category learning because they had shared a value structure during initial learning, albeit a different value structure on different object categories. Thus, performing accurately during the generalization phase of this task provides unambiguous evidence for reliance on an abstract LS representation.

For all blocks of the training and mini-blocked generalization phases, there was no response deadline. The inter-trial interval (ITI) was generated from a lognormal distribution with a mean of 1 s, maximum time of 4 s, and minimum of 0.5 s. Participants’ choice was underlined and the stimuli remained on screen for a 0.5 s inter-stimulus interval. In all blocks, except the generalization test, feedback was displayed for 1.5 s, after which the trial would end. Participants also briefly saw the total gold they had earned within the block after it had ended. However, as there was no way for participants to link these earnings back to the specific trials within the generalization phase, they could not learn the structure of the task from this kind of feedback once they had reached this section of the experiment.

As already noted, this experiment took place over three sessions. The first session was entirely behavioral and served to test how well participants could complete the generalization phase. Prior behavioral pilot experiments had found that approximately 50% of participants did not successfully infer all new context-category values during generalization based on only one experience with training, and so failed to completely learn the latent task structure. As our pre-registered analysis plan depended on participants understanding this structure, we planned, a priori, to exclude from fMRI any participants who did not meet a performance criterion of ≥70% accuracy in all 18 conditions of the generalization test in this first session. Instead these participants were asked to return and complete the rest of the experiment in two additional behavioral sessions. This allowed us to test whether these participants could generalize in this task, given sufficient experience (see main text). That observation helps limit any concerns about generalizability raised by our selection procedures.

In the second and third sessions, after completing the blocked generalization test, participants completed three functional runs of the generalization task but without the nested trial structure (i.e. mini-blocks nested by contexts). These blocks also had a three second response deadline for each trial. In each run, 10 trials from each of the 18 conditions were presented in a pseudo-random sequence optimized for efficiency using Optseq2 with ITIs ranging from 1 to 9 s (180 trials in total) (Greve, 2002). This also ensured that the rate of transitions between trials that shared features was effectively at chance (i.e. 33% for LSs, 16.7% for contexts). After each response, the chosen option (‘sell’ or ‘pass’) was underlined and the stimulus would remain on-screen until the end of the response deadline. A random subset of six of these 12 optimized sequences was selected for the six functional runs for each participant. Thus, each participant completed 1080 trials of this generalization task over the course of two sessions in the scanner (60 trials per 18 conditions), or behaviorally for those participants who did not pass the generalization criterion in the first session. This dense sampling per participant provided more statistical power for our study by reducing within-subject measurement error. Before completing this task, participants also completed a short practice task with the same mixed trial structure with just one trial from each condition.

Participants were given ample information about the structure of the task in instructions that preceded each new phase of the experiment, but not specific information that about which categories were rewarding in which contexts, and which contexts were part of the same kingdoms (i.e. LSs). Specifically, participants were informed that the outcomes were stable but probabilistic (without being explicitly informed of this probability), and were told how many categories of photos each kingdom preferred in each phase of the task. Participants were also instructed that they would later complete a test phase without feedback, so they should strive to commit the values of photo categories in each context to memory. They were also informed of the latent task structure in a general sense, namely that contexts represented locations within distinct kingdoms, and that the relation of contexts to different kingdoms could be uncovered by their shared values for the item categories, but were never explicitly told which contexts were associated with which kingdoms. Participants were also informed that the contexts within each kingdom could appear to be visually very different from each other to discourage them from using such a strategy to link contexts together. To facilitate forming such an abstract task representation, the instructions specifically suggested that participants use a semantic elaboration strategy that tied together the distinct contexts within each kingdom, and the values of the categories within these contexts. These instructions and the cover story about the goblin kingdoms were included to help encourage participants to learn the abstract structure of the task rather than try and individually commit the values of all 54 context-category conjunctions to memory.

At the end of the third session, outside of the scanner, participants completed a similarity judgment task for the context natural scene images that were shown in their specific generalization test. Participants were shown these images on a black background in random starting positions and asked to use the mouse to drag and drop these images on-screen so that their distances reflected the similarity. Participants were specifically instructed not to use any information about to which kingdoms the contexts belonged, but only the visual content of the image. The Euclidean distances between these images were then used to estimate a subjective, participant-specific estimate of the visual similarity of these images used as a nuisance regressor for RSA analyses.

Fifty participants were recruited for this study. Four participants were found to not meet eligibility requirements for the study after session 1 (e.g. they had piercings or implants that were not compatible with MRI), and their data was not analyzed. Twenty-four did not reach criterion performance in the generalization phase (see above) in session 1 and were invited to complete the second and third sessions in behavioral testing settings rather than in the scanner. Six of these participants dropped out of the study before completing both remaining sessions and two did not have complete data due to computer errors. Twenty-two participants passed the generalization criterion and were asked to complete the remaining two sessions in the scanner. Of these participants, two had low accuracy in the generalization phase in the fMRI sessions, indicating a failure to learn the task structure, and their data was not analyzed. Two further participants were excluded because of excessive movement (more than one voxel) in more than one run. One participant had excessive movement only in the last run of the last session and this single run was discarded from analysis. One participant could not complete the experiment due to problems with the scanner on the day of testing, and one participant dropped out of the study before finishing both scanned sessions. In total, 16 participants (10 female, mean age 21.3 years, SD = 3.3 years) passed the generalization criterion on the first day, completed both days of the fMRI experiment and were included in analyses, while 16 participants (11 female, mean age 22.3 years, SD = 2.5 years) did not pass this criterion on the first day and completed the remaining 2 days of testing in behavioral sessions. All participants gave their written informed consent to participate in this study, as approved by the Human Research Protections Office at Brown University, and were compensated for their participation.

The sample size for this study was determined by the target sample size for the fMRI experiment. Data collection was ceased once 16 participants had completed both fMRI scanning sessions while meeting eligibility for inclusion in our analysis (i.e. very little movement and high accuracy during the generalization phase of the experiment). The target sample size was based on a related experiment that successfully used RSA within an OFC ROI (Chan et al., 2016). This target sample was halved, as the current experiment involved two sessions of fMRI scanning and we expected within-subject measurement error to be reduced by this dense sampling approach.

Eighteen images of natural scenes from Konkle et al., 2010 were used to represent contexts within kingdoms (nine for session 1 and nine for sessions 2 and 3). These scenes were chosen to be distinct from each other in content and did not include any visible animals, people, or man-made objects (i.e. the image categories included in the fMRI task). Over the course of the experiment, participants saw 360 images in the initial category training for each image category (randomly sampled from a larger set of 468 images). Hand images were taken from the 11 k Hands Database (Afifi, 2019), leaf images from the Leafsnap database (Kumar et al., 2012), and food images from the Bank of Standardized Stimuli (BOSS) (Brodeur et al., 2014), as well as the Foodpics database (Blechert et al., 2014). In the new category training and generalization phases, 546 images were used from each stimulus category. These included faces from the Chicago Face Database (Ma et al., 2015), animals from the BOSS and CARE databases (Russo et al., 2018), and man-made objects, also from the BOSS database. All category images were cropped or padded with white pixels to fit within a square image with a white background.

To test the rate of change in RTs and accuracy, we fit an exponential function to these data in the form of

where $y$ indicates the predicted trial RT or mean accuracy, $b$ indicates the trial number, and $a$ and $x$ are free parameters that determine the scale and rate of the function respectively. The MATLAB function fminunc (Mathworks, Natick, MA) was used to find values for $a$ and $x$ that produced a function that best fit individual participants’ data using a least squares cost function. This model-fitting process was run 10 times for each parameter fit for each subject, each time with a different random starting point drawn from a standard normal distribution. We chose the parameters that resulted in the lowest sum of squares from these 10 iterations to avoid fits that had converged to local minima.

We developed a set of simple learning models to compare and contrast different explanations for how participants carried out the transfer of option values during the initial generalization phase. The models express retrieval and generalization in terms of spreading activation among a set of nodes that represent the various task elements. In each model, we structured the memory network differently according to the hypothesis tested. However, these models relied on the same basic assumptions and parameters: 1. RTs are a function of the activation strength of the nodes containing retrieved items (Anderson, 1983). 2. The activation strength for each item is initialized at zero and increases according to a Rescorla-Wagner learning rule (Rescorla and Wagner, 1972), controlled by a single rate parameter bounded by 0 to 1 (α₁), each time an option is retrieved, until it reaches an asymptotic value of 1. 3. Activation spreads to contexts and categories linked with retrieved items, controlled by a ‘mediated retrieval’ parameter (α₂). 4. Memory search increases activation in all other task-related items in each trial according to an ‘incidental retrieval’ parameter (α₃). 5. Increases in activation that are inherited from connected nodes falls off according to a power law where the rates of successive steps are multiplied together.

We tested four model variants that each differed in how they represented this memory network, and how activation of task elements propagates during generalization. Importantly, these models are not meant to simulate any particular neural system, but only describe how the format of these representations in memory might differentially affect the retrieval process and, by extension, RTs:

1. CAR – This model implements generalization through associative retrieval by assuming that the memory system stores conjunctions of task-relevant features which can be retrieved if any one of these features is later encountered. This memory network was represented as a matrix of contexts-by-category conjunctions. On each generalization trial, the presented context (e.g. A1) would activate related context-category-value conjunctions from the initial training based on the rate parameter α₁ (e.g. A1-leaves- [negative value], A1-food-, A1-hands+ [positive value]). This would lead to the mediated activation of context-category conjunctions with the same value associations from initial training with the rate parameter α₂ (e.g. A3-hands+, A2-hands+, etc.), which would in turn lead to activation of conjunctions of new categories associated with the context that appeared in the second training phase (e.g. A3-objects-, A3-Faces+, etc.), as determined by the rate of α₂². At the same time, the category presented on the current trial (e.g. objects) would activate conjunctions with that category from the new category training phase (e.g. A3-objects-, B3-objects+) at a rate determined by α₁ The conjunction shown on the current trial (e.g. A1-objects-) would inherit activation from the related conjunction from the new category training phase (e.g. A3-objects-) and increase in activation at a rate determined by α₁²*α₂². Similarly, the same category conjunction with the held out related context (e.g. A2-objects-) would receive a mediated increase in activation at a rate of α₁^2*α₂³. All other nodes not activated by the current trial would also increase their activation as a rate of α₃.

2. IAR – This model also implements generalization through associative retrieval, but rather than relying on episodic conjunctions of context and category, it stores each individual task element independently and links them together based on experience. The memory network was represented as two vectors of contexts and category-value conjunctions (i.e. separate nodes for faces+ faces-, etc.). Each context was linked to their associated category-values from the initial training phase, and only the held-out contexts were directly linked to category-values from the new category training. In each trial (e.g. A1-objects-), activation in the current context (e.g. A1) would increase with the rate parameter α₁. This activation would result in direct retrieval of category-values from initial training at a rate of α₁², and direct retrieval of the context from the new category training phase with the same category-value associations (e.g. A3) at the rate α₁³. Retrieval of this context then led to activation of the relevant category-value node for the current trial (e.g. objects-) from the new training phase at a rate of α₁⁴, allowing generalization, as well as mediated retrieval of other category-values (e.g. faces+, animals+) at a rate of α₁³*α₂. Activation of category values from the initial training would also lead to mediated retrieval of the other related context (e.g. A2) at a rate of α₁²*α₂. All other nodes would increase their activation with a rate set by α₃.

3. LS – This model implements generalization through the formation of a LS that replaces separate representations of individual contexts. Here the memory network was represented as a vector of LSs and a matrix of LSs-by-categories, so category-values were separately nested in each LS. In this case, there was no separate representation of a context, so any increase in activation from a trial involving context A1 immediately carried over to context A2. On each trial, activation in the current LS (e.g. LS A) would increase according to the rate parameter α₁, and in the current category-value node according to α₁². All other category nodes linked to the active LS increase in activation according to α₂, and all other LSs increase activation according to α₃, and their nested categories according to α₃².

4. HLSs – This model also implements generalization through a LS, but builds on the basic LS model to hierarchically cluster contexts within LSs. Thus, this model builds on the LS model but includes a distinct vector for contexts, as well as a matrix of contexts-by-categories. These contexts and context-category nodes are hierarchically nested within each LS (e.g. contexts A1 and A2 are nested within LS A). Thus, activation for each of these contexts and category nodes is inherited from activation of the LSs and LS-category nodes. On each trial, activation in the current context increases at a rate of α₁², and in the current context-category at α₁³. Contexts linked to the current LS also increase in activation according to α₁*α₂, as these contexts inherit mediated activation from the LS. Categories nested in these contexts increase activation according to the rate α₁*α₂². Activation in all other LSs increases at a rate of α₃, contexts not linked to the current LS increase at a rate of α₃², and their nested categories according to α₃³.

For each model, predicted RTs were calculated simply as the sum of the activation level in the current context (or LS in case of the third model) and category, subtracted from the maximum asymptotic activation level for both combined. These values were then z-scored and compared to participants’ z-scored RTs so that both model predictions and behavioral data were on the same scale. The fit of each of these models was calculated as the negative log-likelihood for a simple linear function. Only correct responses were included in this analysis, as the model assumes correct recovery of option values from memory on each trial. Across all 16 participants in the fMRI group, only 17 trials were errors in session 2, and thus accounted for a very small fraction of the data (0.6%). Parameters were optimized using the MATLAB function fmincon, with each model fit for each participant 30 times with random starting points in order to avoid convergence on local minima.

To test whether these models made substantially different RT predictions, we carried out a cross-model comparison using simulated data. We simulated data using the optimized parameters from each participant for each model. We then fit each model on these simulated data, in the same way as with participants’ behavioral data and calculated the negative log-likelihood of this fit to test if the generative model provided a better explanation of the simulated data than the three alternatives.

Whole-brain imaging was acquired using a Siemens 3T Magnetom Prisma system with a 64-channel head coil. In each fMRI session, a high resolution T1 weighted MPRAGE image was acquired for visualization (repetition time (TR), 1900 ms; echo time (TE), 3.02 ms; flip angle, 9°; 160 sagittal slices; 1 × 1 × 1 mm voxels). Functional volumes were acquired using a gradient-echo echo planar sequence (TR, 2000 ms; TE, 25 ms; flip angle 90°; 40 interleaved axial slices tilted approximately 30° from the AC-PC plane; 3 × 3 × 3 mm voxels). Functional data were acquired over three runs. Each run lasted 15.1 min on average (452 acquisitions). After the functional runs, a brief in-plane anatomical T1 image was collected, which was used to define a brain mask that respected the contours of the brain and the space of the functional runs (TR, 350 ms; TE 2.5 ms; flip angle 70°; 40 axial slices; 1.5 × 1.5 × 3 mm voxels). The sequence of scans was identical on both sessions. Soft padding was used to restrict head motion throughout the experiment. Stimuli were presented on a 32-inch monitor at the back of the bore of the magnet, and participants viewed the screen through a mirror attached to the head coil. Participants used a five-button fiber optic response pad to interact with the experiment (Current Designs, Philadelphia, PA).

Functional data were preprocessed using SPM12. Quality assurance for the functional data of each participant was first assessed through visual inspection and TSdiffAna (https://sourceforge.net/projects/spmtools/) and ArtRepair (http://cibsr.stanford.edu/tools/human-brain-project/artrepair-software.html). Outlier volumes (defined by standard deviations from the global average signal) were interpolated when possible. If interpolation was not possible, a nuisance regressor was added to the run with a stick function at the time points for these volumes. Slice timing correction was carried out by resampling slices to match the first slice. Next, motion during functional runs and days was corrected by registering volumes to the first volume in the first session using rigid-body transformation.

A deformation matrix for spatial normalization to Montreal Neurological Institute (MNI) stereotaxic space using fourth order B-spline interpolation was calculated for the motion corrected functional volumes. The in-plane T1 anatomical image was used to create a brain mask for functional analysis using the Brain Extraction Tool in FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/). This mask was then normalized to MNI space using the inverse deformation matrix from the normalization of the functional data.

Functional data were analyzed under the assumptions of the GLM using SPM12. Separate regressors were included for correct, erroneous, and missed responses for each condition with the duration of each response set to participants trial-wise RT (or the duration of the stimulus display for missed responses). Nuisance regressors for participant motion (six translational and rotational components) were also included, as was an additional regressor for scan session. Regressors and parametric modulators were convolved with the SPM canonical hemodynamic response function (HRF). Functional data were pre-whitened and high-pass filtered at 0.008 Hz.

Whole-brain searchlight and ROI-based RSA were carried out in a two-step process. 

First, in participant-level analyses, an empirical RDM was estimated from the cross-validated Mahalanobis distances of the regressors for the 18 conditions (six contexts × three categories) averaged over six runs within the generalization phase from a designated subset of voxels in each participant (either defined by an ROI or a spherical searchlight) using the RSAtoolbox v.2.0 (Nili et al., 2014; Walther et al., 2016). The lower triangle of this empirical RDM was extracted giving the distances for the 153 condition pairs in this multivoxel space.

Hypothesis RDMs were constructed based on the similarities/dissimilarities that would be expected for a pure representation of a given type. These included five main effects (LS, context, value, image category, and subjective visual similarity of contexts), as well as five interactions (LS × category, LS × value, value × category, context × category, context × value).

In the case of LS, value, and context RDMs, the hypothesized distances were simply set so that conditions that were the same on these factors would have smaller expected distances and larger distances where these conditions were not the same. For the category RDM, hypothesized distances were based on asymmetric distances related to the perceptual-semantic distances of animate and inanimate stimuli observed in many other studies (Chikazoe et al., 2014; Thorat et al., 2019). Namely, that animate categories (faces and animals) were expected to be more similar to each other than inanimate objects, but inanimate objects were expected to be more similar to animals than to faces. Distances for these hypothesis RDMs were set as ordinal integer numbers to reflect predicted distances (e.g. for the category RDM, the expected distance between two conditions with faces would be 1, the distance between faces and animals would be 2, and between faces and objects would be 4). The RDM for the subjective visual similarity of contexts was specific to each participant and derived from the Euclidean distances of these natural scene images in the similarity judgment task completed at the end of the third session of the experiment. Interaction RDMs were created by extracting main effect RDMs below the diagonal (i.e. the lower triangle of the RDM matrix) as a vector, z-scoring these values and multiplying them together. These interaction regressors allowed us to test where pattern-similarity reflected the modulation of one task component another (e.g. where items belonging to different categories were represented more similarly because of shared value associations).

Second, these hypothesis RDMs were related to the empirical RDM through multiple linear regression analysis, where a coefficient was estimated relating each hypothesis RDM to empirical RDMs allowing us to parcel out variance in representational distances due to multiple factors in the same model (e.g. Nassar et al., 2019). Main effects and interaction terms within this model were allowed to compete for variance simultaneously. The lower-triangle of all hypothesis RDMs were extracted as vectors, z-scored, and included as predictors, along with an intercept term, in a multiple linear regression analysis to calculate beta coefficients relating each hypothesis RDM to the empirical RDM.

Whole-brain analyses were carried out by passing a spherical searchlight with a radius of 9 mm over each voxel within participants’ brain mask in native space. For each participant, beta coefficients for hypothesis RDMs were calculated at each step and averaged over searchlight passes for all voxels included in the searchlight to compute fixed-effects in a first-level analysis. This approach is similar to a common approach of assigning coefficients to a the central voxel of each searchlight and then smoothing these maps before second-level tests (e.g. Devereux et al., 2013), but requires one less experimenter degree of freedom in defining the full-width half-maximum (FWHM) of the smoothing kernel. Group level analyses were conducted by normalizing participants’ beta coefficient maps to MNI space using the deformation field from the normalization of participants’ functional data and computing a one-sample t-test against zero. These volumes were kept in a 1.5 × 1.5 × 1.5 mm space and not resampled. Whole brain t-statistic maps were thresholded at a cluster defining threshold of p<0.001 uncorrected. Non-parametric permutation tests (10,000 permutations) were used to derive a cluster extent threshold (k) for each test by randomly flipping the sign for half of participants’ contrasts and generating a null distribution based on the suprathreshold maximum cluster statistics. The cluster extent threshold in each contrast was defined as the 95th percentile of this null distribution in order to test for statistical significance at p<0.05, corrected for multiple comparisons. (Nichols and Holmes, 2002).

We defined three a priori anatomical ROIs in this study based on prior work using the Automated Anatomical Labelling (AAL2) atlas (Rolls et al., 2015). First, the OFC ROI was given the same definition used by a study examining LS representations in this region (Schuck et al., 2016), which followed that of Kahnt et al., 2012 (Kahnt et al., 2012). This definition included the bilateral combination of the following regions: the superior orbital gyri, middle orbital gyri, inferior orbital gyri medial orbital gyri, and rectal gyri. The VTC ROI was given the same definition used by a study that showed that pattern activity in this region differentiated between visual images based on animacy (Chikazoe et al., 2014), excepting the bilateral parahippocampal gyri. This definition included the bilateral lingual gyri, fusiform gyri, and inferior temporal cortices. The HPC ROI was defined as the bilateral hippocampi. These masks were warped into participants’ native brain space using the inverse deformation matrix for all ROI-based analyses. We first conducted an RSA analysis using all voxels within these ROIs by calculating empirical RDMs for the cross-validated Mahalanobis distances from voxels within these ROIs in the same way as in the searchlight analyses, and then using the same multiple linear regression model to relate hypothesis RDMs to these empirical RDMs. These beta-coefficients were then subjected to a second-level, one-sample t-test against zero to estimate statistical significance.

These ROIs were also used as an explicit mask in searchlight analyses to test for effects that were below threshold at a whole-brain cluster-corrected level within HPC and OFC. For these analyses, statistical significance was evaluated as in the whole-brain searchlight analysis with cluster-based permutation tests to compute a cluster extent threshold within these smaller volumes for each contrast, controlling for multiple comparisons at p<0.05.

Functional volumes were normalized to a 1.5 × 1.5 × 1.5 mm MNI space and smoothed with an 8 mm FWHM Gaussian isotropic kernel. Beta coefficients for single subject effects were estimated using a fixed-effects model in a first-level analysis. For whole-brain contrasts, these estimates were then included in a second-level analysis using a one-sample t-test against zero at each voxel. As with whole-brain RSA analyses, t-statistical maps were thresholded at a cluster forming threshold of p<0.001 and cluster-based permutation tests were used to compute a cluster extent threshold controlling for multiple comparisons at p<0.05.

The results of the whole-brain LS RSA searchlight was compared with a cortical parcellation based on resting state functional connectivity data from 1000 individuals by Yeo et al., 2011. To find the degree of overlap within each of these functional networks, we calculated the proportion of voxels in the cluster-corrected LS statistical map that fell within these 17 networks projected to MNI152 space and defined with liberal boundaries around the cortex.

To assess the stability of task-relevant representations across scanning sessions, we carried out two separate searchlight RSA analyses using data from within each session, as above. The resultant coefficient maps from the multiple linear regression analysis were then contrasted between sessions in both directions using paired t-tests. These contrasts were then assessed with the same cluster-forming threshold and cluster-based permutation tests to control for multiple comparisons, as in univariate and multivariate analyses.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

This work was supported by a Multidisciplinary University Research Initiative award from the Office of Naval Research (N00014-16-1-2832), a Postbaccalaureate Research Education Program grant from the National Institute of General Medical Sciences (R25GM125500), and a Ruth L Kirschstein National Research Service Award to AV from the National Institute of Mental Health (F32MH116592). We thank Apoorva Bhandari and other members of the Badre Lab for constructive discussion and consultation on task design and fMRI analysis, as well as Linda Q Yu for comments on this manuscript. We also thank Emily Chicklis for assistance with data collection.

Human subjects: All participants gave their written informed consent to participate in this study, as approved by the Human Research Protections Office at Brown University, and were compensated for their participation.

© 2021, Vaidya et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.