(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Neural traces of composite tasks in complex task representation in the human brain reflects learning performance [1]

['Woo-Tek Lee', 'Department Of Psychological', 'Brain Sciences', 'University Of Iowa', 'Iowa City', 'Iowa', 'United States Of America', 'Cognitive Control Collaborative', 'Behavioral-Biomedical Interface Training Program', 'Eliot Hazeltine']

Date: 2026-01 

Task knowledge can be encoded hierarchically such that complex tasks can be built by associating simpler tasks. This associative organization supports generalization to facilitate learning of related but novel complex tasks. To study how the brain implements generalization in hierarchical task learning, we trained human participants on two complex tasks that shared a simple task and tested them on novel complex tasks whose association could be inferred via the shared simple task. Behaviorally, we observed faster learning of the novel complex tasks than control tasks. Using electroencephalogram (EEG) data, we decoded constituent simple tasks when performing a complex task (i.e., EEG association effect). Crucially, the shared simple task, although not part of the novel complex task, could be reliably decoded from the novel complex task. This decoding strength was correlated with the EEG association effect and the behavioral generalization effect. The findings demonstrate how task learning can be accelerated by associative inference.

During performance, we obtained high temporal resolution and high-density electroencephalogram (EEG) data, to which we applied decoding and representational similarity analysis (RSA, [ 10 , 33 ]). To preview the results, the behavioral data revealed faster learning of complex tasks consisting of indirectly associated simple tasks compared to control tasks consisting of equally practiced simple tasks that were not indirectly associated. This result replicates the behavioral generalization effect reported in [ 6 ] and suggests that participants learned feature-based tasks efficiently by optimizing cognitive control mechanisms. We then observed robust decoding of constituent simple task representations from the complex task EEG data, which validates that our analysis is sufficient to capture dynamics of abstract task representations. Critically, we can decode the latent task (e.g., B) that is reinstated by the indirect association in the complex task (e.g., AC) EEG data. We further found that the strength of the latent task representation in complex task EEG data is linked to both the strength of the constituent simple task representations in complex task EEG data and generalization performance. In sum, the results suggest that complex task learning results in a network representing associated information that enables generalization via the reinstatement of the latent task. The strength of generalization is linked to the strength of the associations.

To test these predictions, we leveraged a behavioral paradigm of generalization in task learning [ 6 ]. In this paradigm, subjects performed tasks that require to attend task-relevant feature(s) based on the cue(s). While the ‘task’ in this paradigm is strongly intertwined with feature-based attention, our use of the term is consistent with classic task-switching literature [ 28 – 30 ], where a ‘task’ is requires the process that enables the efficient guidance of attention toward task-relevant feature. Furthermore, the task-switching literature provides convincing evidence that a ‘task’ cannot be fully explained by feature-based attention [ 29 – 32 ]. For instance, mixing costs, which refer to the cognitive cost observed when performing a task in a mixed-task block (i.e., on each trial one of multiple possible tasks is performed) compared to a single-task block, cannot be fully explained by shift of feature-based attention, since the cost occurs even when participants repeat the same task, implying no change in feature-based attention [ 29 , 30 ].

Applying this idea to the context of task learning, we predict that the reinstatement of the latent task from the complex task is expected to be key for generalization, which relies on the associations between simple task representations. Moreover, both theories predict that stronger latent task reactivation will lead to more efficient learning, as the reactivation supports the inference generated by indirect associations between constituent task representations via stronger activation of integrated representations [ 20 , 22 , 25 ] or stronger reinstatement of constituent representations that are connected with latent representation [ 21 ]. However, while the reinstatement of memory items for generalization has been extensively studied, whether the same neural mechanism applies to abstract task representations, and if so, how the latent representations support efficient task learning, has not been directly tested. Here, we predict that the co-activation of simple task representations during complex task learning leads to associations between simple tasks and that these associations can facilitate the new complex learning consisting of indirectly associated simple tasks. Based on these predictions, we hypothesize that the generalization will lead to reinstatement of latent task representation. Additionally, we hypothesize that stronger neural evidence of associative activation will predict stronger generalization.

While these two major theories differ in terms of their proposed mechanisms of generalization, they share an important prediction about how the generalization occurs between associated memory representations via associative inference. When there are multiple associated item pairs that share a component (e.g., A-B and B-C association), activation of nonoverlapping items (i.e., A and C) should reinstate the latent (i.e., not presented; B) item, because of the pattern completion processes of integrated representation (integrative encoding; [ 22 ]), or feedback signals from conjunctive representations that are linked with activated items (recurrent interaction; [ 21 ]).

Two prominent theories of association and generalization in episodic memory offer insights into the processes supporting abstract task learning and generalization. Integrative encoding theory [ 22 ] posits that repeated co-activation of items (e.g., A and C) will create an associated memory representation (e.g., AC). Moreover, integration can occur between associated item pairs that share an item (e.g., AB and BC), which leads to an integrated representation (e.g., ABC) that encompasses all three memory items [ 20 , 22 , 25 ]. The integrated representation will be retrieved if a subset of the items is present. For example, presentation of AC is expected to reinstate item B as well as items A and C. On the other hand, recurrent interaction theory [ 21 ] proposes that an association between two memory items forms a conjunctive representation, and these conjunctive representations are connected to each of the constituent memory items. If a representation is activated (e.g., by sensory input), its connected representations will also be activated (though to a lesser extent). For example, if two conjunctions AB and BC share a memory item B, activating the conjunction AB will also activate memory items A and B. B will further activate conjunction BC, which in turn activates C. In the end, the co-activation of A and C will facilitate the formation of an indirect association of AC.

This associative feature of task representations allows complex tasks to share components, which enables humans to generalize from previous experience and quickly adapt to new, related tasks [ 15 ]. For example, humans can extract structure from experiences and infer novel knowledge from the structure, a phenomenon sometimes also termed zero-shot learning [ 16 – 18 ]. While generalization can occur via different mechanisms, such as rule-based or similarity-based [ 19 ], generalization through association requires additional inferential processes as it allows for learning without direct association or experience. This generalization in task learning may be attributed to associative inference, which refers to the ability to infer novel, indirect associations from partially overlapping associations [ 9 , 11 , 20 – 27 ]. For example, the barista who is skilled at the ‘café latte’ task (espresso + milk) and the ‘café mocha’ task (espresso + chocolate) can infer how to make a hot chocolate via the ‘making espresso’ task shared by the ‘café latte’ and ‘café mocha’ tasks. In the laboratory, we reported that generalization via associative inference accelerates task learning and improves task performance [ 6 ]. Although generalization by associative inference plays an important role in task learning [ 3 , 6 , 9 ], how it is implemented in the brain is understudied.

Humans have the remarkable ability to efficiently learn new complex and cognitively demanding tasks. To do this, we organize task representations hierarchically [ 1 – 4 ], such that complex tasks can be represented by associating simpler ones. That is, new complex tasks can be learned efficiently by associating existing simple task representations and inferring their direct and indirect connections [ 5 , 6 ]. For example, a barista can learn how to make a ‘café latte’ by associating the existing skills of making espresso and steaming milk. This ability is analogous to studies of human memory, where researchers have found that associating multiple memory items enables us to adaptively adjust behavior by predicting future events [ 7 ] or inferring novel information [ 8 ]. This associative feature of memory representations has been extended to incorporate the cognitive control parameters and task demands [ 5 , 6 , 9 – 13 ] that support adaptive, goal-directed behavior. Although the lateral prefrontal cortex (lPFC) has been shown to be involved in representing task rules for goal-directed behavior [ 14 ], little work has addressed how simple tasks can be associated and generalized for building of complex task representations.

Results

Experimental design Participants (n = 40) first learned six simple tasks (coded as A-F), each of which was a feature-based attention task defined on one out of the six stimulus features (Fig 1A). On each trial, participants judged whether a cued feature appeared in a forthcoming target stimulus (Fig 1B). After the simple task phase, participants learned complex tasks, each consisting of two simple tasks. On each trial of complex task, participants followed the ‘exclusive or’ design, judging whether or not only one of two cued features matched the target stimulus. Because there were two possible responses, the chance level of correct response was 50%. All feature values were presented equally frequently during the complex task phase. During the training stage of the complex task phase, different complex tasks, denoted as the pair of constituent simple tasks (e.g., AB means a complex task consisting of simple tasks A and B), were presented on alternate blocks (odd blocks: AB, DE, even blocks: BC, EF; Fig 1C). In the training stage, there were two sets of complex tasks, such that each set contained two complex tasks sharing one simple task (e.g., set 1: AB and BC, set 2: DE and EF). After the training stage, participants performed four new complex tasks (AC, DF, AD, and CF) in the test stage (Fig 1C). Two of the new complex tasks could be generalized via associative inference because both simple tasks were associated with the same simple task in practiced complex tasks (e.g., AC from AB and BC; Fig 1D) and are hence termed generalizable tasks. In contrast, the other two new complex tasks (e.g., AD and CF) could not be generalized using associative inference, as the constituent simple tasks did not overlap during the training stage. These were termed control tasks. Generalizable and control complex tasks were presented equally frequently within each block in random order. Within this behavioral framework, we defined task representation as cognitive processes involving task-relevant feature information and further posited that a major component of learning involves the improved application of cognitive control and associative mechanisms, leading to optimized feature-based selection and processing and enhanced behavioral performance, based on previous research [6]. This design enabled us to investigate the behavioral generalization effect by comparing performance between generalizable and control tasks, while also examining the temporal dynamics of associative and generalization effects through EEG data. We note that the assignment of each simple task to a specific label (e.g., A, B, C, D, E, or F) was always random for each participant. For additional confirmation, we visualized frequencies of each simple task assigned to each specific label (S6 Fig). This histogram showed no systematic bias of simple task assignments. This result supports that the reported findings are not mainly driven by a particular simple task. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Experimental design and behavioral results. (A) Cue and target stimuli. Each cue indicated one stimulus feature and was presented as either text or image. Two example stimuli are illustrated in the rightmost column. (B) Trial events of the simple and complex tasks. In the simple task, participants first encountered a single cue that predicts a feature of an upcoming stimulus. After a delay of 500 ms, the target stimulus was presented until response or a deadline. Participants judged if the cue matched the target with button press. Visual feedback was provided after another 500 ms delay. The complex task is identical to the simple task except for two changes. First, participants encountered two cues instead of one. Second, participants decided whether only one cue matched the target (i.e., the ‘exclusive or’ rule). (C) Illustration of the experimental design. (D) Predicted dynamics of task representations in the complex task phase. Participants were expected to associate simple task representations to build a complex task representation and to further generalize two complex tasks via associative inference. Thus, we predicted that the generalizable complex tasks present faster decrements in RT than the control complex tasks. (E) Block-wise RTs (left, group mean and standard error of means, SEM) and slope of RTs (right, group mean and SEM, dots represent individual slopes) of complex tasks for each condition during the test stage. The data and code needed to generate this figure can be found at https://doi.org/10.17605/OSF.IO/MZF4A. https://doi.org/10.1371/journal.pbio.3003613.g001

Behavioral data Group-average accuracy for both simple and complex tasks stayed above 80% throughout the experiment (S1 and S2 Figs). Performance varied significantly across individual simple tasks, particularly for the ‘fill’ task (S1 Table), so we conducted additional analyses to address potential confounding effects in the EEG data analysis driven by these behavioral differences, as detailed in the ‘Above-Chance Decoding in Simple Task EEG Data’ section. To assess the generalization effect, we compared the performance of the generalizable and control conditions in the test stage (Table 1). Accuracy for the generalization and the control conditions did not differ significantly (t(39) = 1.56, p = .126, Cohen’s d = .25). For the RT analysis, we followed our previous procedure [6] and regressed out confounding factors such as performance differences among simple tasks, response switch costs and so on (see Behavioral analysis in Materials and methods for details). To control potential high multi-collinearity between confound regressors, we calculated variance inflation factor (VIF) scores for each regressor and confirmed that all VIF scores are lower than 2 (S2 Table). The resulting residual RTs were used to test the hypothesis of generalization. Replicating our previous finding [6], we found significantly faster decrease in RT over time (t(39) = −2.03, p = .049, Cohen’s d = .32) in the generalizable complex tasks than in the control complex tasks (Fig 1E), indicating faster learning for generalizable than control complex tasks. To account for the nested design of the present experiment, we performed a linear mixed-effect model analysis and found a consistent significant effect of learning speed (t(4398.14) = −2.10, p = .036). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 1. Behavioral results of the test stage in complex task phase. Group mean (SEM) of accuracy and residual RT in each condition (Gen = generalizable, Con = control) and block of the test stage. https://doi.org/10.1371/journal.pbio.3003613.t001

Decoding simple task EEG data For each participant and each time point (i.e., a 50 ms bin) of the simple task phase (Fig 1C), we trained a multiclass decoder on the EEG data to classify the six simple tasks. The decoding analysis was performed after the onset of the target stimulus (Fig 1B), so the decoder performance was not affected by perceptual processes related to the cue. Cross-validation decoding accuracy of simple tasks increased shortly after the target onset and was significantly above-chance after 50 ms, peaking at around 300 ms (Fig 2A, left panel). The longest epoch that showed significant decoding performance (indicated in red line in left panel of Fig 2A) survived nonparametric permutation test for multiple-comparison correction (p corrected < .001). The topographical map of the averaged pattern matrix [34] shows strong absolute weights in occipital and posterior electrodes, with notable contributions from left frontal electrodes, suggesting both sensory and cognitive processes (Fig 2A, right panel). This result suggests that the simple task representations can be robustly decoded in EEG data after target stimulus onset during their performance. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. EEG decoding results. (A) Simple task EEG data decoding results. Left: decoding accuracy over time within a trial. The gray shade indicates SEM of the decoding accuracy. The blue line indicates chance-level decoding accuracy. The red line indicates time points showing significantly above-chance decoding accuracy. Right: visualization of mean electrode-wise contributions to the decoders based on pattern matrix [34]. (B) and (C) EEG association and generalization effects of complex task EEG data. In each panel, decoding performance (averaged t-values from trial-level decoding analysis, see Materials and methods) is plotted as a function of the time point of training (Y-axis) and the time point of test (X-axis). (B) T-values indicating decoding strength of constituent simple task representations from complex tasks in the training stage (left panel; e.g., A and B in AB) or in the test stage (right panel; e.g., A and C from AC). (C) T-values indicating decoding strength of the latent simple task in the training stage (left panel; e.g., C in AB), or in the test stage compared to nonassociated latent simple task (right panel; e.g., B–E in AC). The time points are time-locked to the stimulus onset. (D) T-values indicating decoding strength of post-hoc decoding analysis testing above-chance decoding of the latent task (e.g., B in AC). (E) Results of decoding the latent tasks in control complex task (e.g., B and E in AD). The data and code needed to generate this figure can be found at https://doi.org/10.17605/OSF.IO/MZF4A. https://doi.org/10.1371/journal.pbio.3003613.g002 Participants were required to perform the simple tasks with the same response deadline (see Materials and methods), which likely contributed to the significantly different behavioral accuracies between simple tasks (S1 Table). To address the potential confounding effect in the decoder’s performance caused by the different number of trials across simple tasks, we calculated decoding accuracies separately for each simple task. Although the decoding accuracies differed among simple tasks, all tasks showed significantly higher decoding accuracies at most time points. More importantly, the decoding accuracies between simple tasks did not correlate with their behavioral accuracies, and the least accurately performed ‘fill’ task showed similar, or higher, decoding accuracies compared to other simple tasks (S3A Fig). The temporal generalization analysis in each task further showed no systematic difference in their temporal dynamics of decoding performance across different simple tasks (S4 Fig). Overall, this result indicates that the observed difference in behavioral performance between simple tasks did not systematically confound the decoders’ performance.

Overview of cross-phase EEG decoding analysis To assess the reinstatement of simple tasks during the performance of complex tasks, we trained decoders on simple task phase EEG data, as described above, then tested on complex task phase data (see Materials and methods), only using correct trials in both simple tasks and complex tasks. The mean (SD) number of trials included in the analysis was 293.48 (19.15) for the simple task phase, 215.05 (20.10) for the complex task training stage, and 52.68 (5.56) for the complex task test stage. To test the association effect (e.g., decoding simple tasks A and C from complex task AC), the decoders were tested on complex task EEG data from the training stage (blocks 1–8). To test the generalization effect (e.g., decoding simple task B from complex task AC), the decoders were tested on the test stage of the complex task phase (blocks 9–11, Fig 1C). For each test time point, the decoder produced a probabilistic prediction of each simple task being represented, which sums up to 1 across all simple tasks in each trial. We logit-transformed the decoding probability data, which prevent issue caused by the fact that sum of all classes’ probability is 1. Next, we regressed the logits against a linear model with predictors representing association effect (i.e., constituent simple task), generalization effect (i.e., the latent simple task), and the baseline of the nonassociated latent task (e.g., E in AC complex tasks; see Representational similarity analysis in Materials and methods for more details). This baseline was included to control for any potential confounding effect in the generalization effect in the test stage due to the higher presentation frequency, since the latent tasks (i.e., B in AC and E in DF) always presented twice as often as other simple tasks during the training stage, which addresses the issue that the activation levels of constituent tasks would affect the baseline level. The trial-wise regression coefficient of each predictor was averaged for each participant and tested against 0 using a one sample t-test. The presence of an association effect and a generalization effect in training stage predicts that beta values will be greater than 0. For the generalization effect in the test stage, we subtracted the effect of the baseline regressor, thus testing the prediction that decoding performance of the associated latent task will be greater than that of the nonassociated latent task (e.g., B minus E in AC complex tasks). We used the t-values to quantify the strength of the association and generalization effects. For the multiple-comparison corrections, we applied a nonparametric permutation test [35].

EEG association effect We first investigated the activation of constituent simple task representations to test the hypothesis that constituent simple task representations would be decodable during complex task trials. The purpose of this analysis was to assess the association between simple task representations for complex task learning from EEG data and to validate the simple task decoder’s performance when applied to the complex task EEG data. The decoding strength of constituent simple tasks can relate to the associations between task representations in two ways. One possibility is the subsequent memory effect [36–40]. In this study, the subsequent memory effect posits that EEG decodability reflects neural representation strength of the constituent tasks, which in turn is linked to the encoding strength of the association between constituent tasks. Alternatively, performing constituent tasks to complete the complex task may form associations between the constituent tasks. The association, when pair with one constituent task as the cue, may reinstate the other constituent task, leading to stronger task representation on top of that caused by executing the other constituent task. Fig 2B shows the time points that the constituent simple task representations can be decoded from the complex task EEG data (e.g., A and B from AB) in the training and test stages, respectively. In the training stage, constituent simple task representations were decoded above-chance level from complex task EEG data between 150 ms and around 1,200 ms post target onset when the decoders were trained between around 150 ms and 1,000 ms post target onset (p corrected < .001; Fig 2B, left panel). In the test stage, this EEG association effect remained significant (p corrected = .041; Fig 2B, right panel) yet was reduced, significant between around 150 ms and 700 ms post target onset in training data and between 150 and 550 ms post target onset in test data. These results indicate that complex task execution involves constituent simple task activation, which is more extensive in the training than in the test stage.

EEG generalization effect While representing complex tasks as associations between simple task representations can be beneficial for their initial learning, generalization can support the learning of a novel complex task if its associations can be inferred from other learned complex tasks. Both integrative encoding theory [22] and recurrent interaction theory [21] predict the reinstatement of the latent task representation during test stage complex task trials (e.g., B from AC), despite the fact that the latent task is not present on the trial. To test this, we applied decoding analysis and RSA to assess the reinstatement of the latent simple task representations in the test stage complex task EEG data. Consistent with the prediction, latent task (e.g., B in AC) representations, compared to nonassociated latent task (e.g., E in AC), were better decoded between 850ms to 1,000 ms post target onset in the test stage using decoders trained between 200 ms and 300 ms post target onset (Fig 2C, right panel, p corrected = .048). While both theories predict that latent task represent should be presented in the test stage, integrative encoding theory further posits that integrative memory can form when participants experience both complex tasks that share a simple task [11,22]; e.g., AB and BC) during the training stage. This suggests that generalization may occur during the training stage, starting from the third block. To test this prediction that generalization occurred during the training stage, we tested the decoding of the latent task in the training stage (e.g., C from AB). In contrast to the test stage data, no time points reached statistical significance after the cluster-based nonparametric permutation test. As a confirmatory analysis, and to investigate temporal dynamics of associated latent task reactivation (e.g., B in AC) compared to unassociated latent task reactivation (e.g., E in AC), we conducted a post-hoc analysis, assessing latent task reactivation without the baseline regressor for the nonassociated latent task (Fig 2D). In this post-hoc analysis, latent task representations were decoded above-chance level between 850 ms and 2,000 ms post target onset in the test stage using decoders trained between 200 ms and 300 ms post target onset (Fig 2, right panel, p corrected = .047). This result supports the main EEG generalization effect by showing overlapping time points between 850 msand 1,000 ms post target onset in the test stage. Furthermore, this result suggests that the latent task reactivation in later time points in complex task EEG data (1,000–2,000 ms) might be confounded by the higher presentation frequency of latent task as supported by the above-chance decoding of the unassociated latent task after 1,000 ms of target onset (S6 Fig). Specifically, latent task in test stage presented twice more often than the other tasks during training stage, which might evoke late sustained activation during the test stage, regardless of whether the latent task that is relevant to the current complex task [41,42]. As the control complex task also contained simple tasks with learned associations (e.g., A-B and D-E in AD complex task), it is possible that these latent tasks (B and E in AD) were reactivated by these associations. To test this, we regressed the logits of decoding performance for each task against a linear model with regressors for association and generalization effects in control complex tasks, and tested decoding performance of the latent simple tasks in the test stage complex task EEG data (Fig 2E). None of the clusters survived after controlling for multiple comparisons (p corrected = .190). Additionally, we analyzed latent task reactivation by averaging its activity within the time window that showed significant EEG generalization effect in our main analysis (Fig 2C, right panel). A t-test performed on these values revealed no significant effect (t(39) = 1.63, p = .110). However, the difference between latent task reactivation in control complex task and generalizable complex task did not show significant effect (t(39) = 0.57, p = .571). The results did not support the reactivation of latent tasks in control complex tasks.

Correlation between EEG association and generalization effects Two major accounts of generalization in episodic memory [21,22] hold that the association effect (i.e., decodability of constituent simple tasks during complex task performance) lays the foundation of generalization effect (i.e., decodability of latent task from complex task in the test stage). Therefore, we predicted that the magnitude of the association effect should correlate with magnitude of the generalization effect. To test this correlation, we first calculated averaged EEG generalization effect for each participant within a time window defined in the following manner (Fig 3A): Based on the EEG generalization effect in Fig 2C and 2E, we calculated the mean EEG generalization effect within the time window that was significant in both main and post-hoc EEG generalization effect: 200–300 ms post target onset for the simple task data and 850–1,000 ms for the complex task data. The decision to use 200–300 ms post target onset for the simple task EEG data is further supported by simple task decoding analysis, which showed the most robust decoding performance in that time window (Figs 2A and S4). For consistency, the EEG association effect was trained on simple task EEG data from the same window (200–300 ms post target onset) and tested on time points showing statistically significant association effect, which starts at 100 ms (Fig 2B, left). The test window was truncated at 850 ms to avoid temporal overlap between the periods used to measure the association and generalization effects. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. Correlation analyses results. (A) Data used for correlation analyses. For each individual, EEG association effect (left panel), EEG generalization effect (center panel), and behavioral generalization effect (right panel) were used. The EEG association effect was calculated by averaging single-trial t-values from the decoders trained on simple task time points between 200 ms and 300 ms and tested on training phase complex task time points between 200 ms and 1,000 ms. The EEG generalization effect was calculated by subject-mean single-trial t-values from the decoders to be trained on simple task time points between 200 ms and 300 ms and tested on test phase complex task time points between 850 ms and 1,000 ms. The behavioral generalization effect was calculated as the difference in RT slope (Fig 1E) between generalizable and control conditions. (B) Scatter plots showing individual EEG association and EEG generalization effects (left) or EEG generalization effect and behavioral generalization effects (right). (C) Scatter plots showing individual EEG generalization effect in block 9 and residual RTs for blocks 9, 10, and 11, presented from left to right. The data and code needed to generate this figure can be found at https://doi.org/10.17605/OSF.IO/MZF4A. https://doi.org/10.1371/journal.pbio.3003613.g003 Consistent with the prediction, EEG association effect in the training stage showed a significant positive correlation with the EEG generalization effect (r(38) = 0.36, p = .024; Fig 3B, left panel). Overall, these results indicate that the association of simple task representations during the training stage of the complex task phase was linked to the EEG generalization effect before and near the response in the test stage.

Correlation between behavioral and neural effects of generalization We expected that the EEG generalization effect would benefit performance of generalizable complex tasks, whether it is through stronger activation of the integrated memory representation with pattern completion process (integrative encoding) or through stronger activation of the conjunctive representation that is connected with latent memory representation [21]. Therefore, we conducted an exploratory analysis to investigate the behavioral relevance of the EEG generalization effect. We first assessed individual behavioral generalization effects using the RT slope difference between generalizable and control conditions (generalizable − control) and correlated this behavioral generalization effect with the EEG generalization effect from the previous analysis (Fig 1E). The slope differences showed a significant negative correlation with the EEG generalization effect (r(38) = −0.32, p = .043), indicating that a stronger EEG generalization effect is linked to larger difference in RT decrease in generalizable tasks compared to control tasks (Fig 3B, right panel). The accuracy difference between generalizable and control complex tasks did not show any significant effect with EEG generalization effect (r(38) = 0.11, p = .506). This result indicates that reactivation of latent task in test stage supports the overall learning speed of generalizable complex tasks.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3003613

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/