Neural representation of abstract task structure during generalization
- Department of Cognitive, Linguistic, and Psychological Sciences, Brown University, United States
- Department of Psychology, Stanford University, Stanford, United States
- Department of Psychology and Brain Sciences, University of Massachusetts Amherst, United States
- Carney Institute for Brain Science, Brown University, United States
Figures
Figure 1 with 1 supplement
Schematic of the experimental task, and its design and logic.
(a) In each trial of the training and generalization phases, participants were asked to make a decision to sell or pass on an image, the value of which depended on contexts shown above the image. Throughout the two training phases, participants received feedback on every trial, but not during generalization. Participants saw three categories of images in the same context over small batches of trials for each unique combination (termed a mini-block) before switching to a new context. (b–d). The left tables show the reward structure for example context-category pairs across the three phases of the experiment in a single session. Cells show the probabilities of reward for each pair. The right schematics illustrate clustering of contexts by category-values into latent states (LSs; blue and orange arrows) and inference of values via structured knowledge (red arrows). Only two LSs are shown for visualization. Latent State (LS) Hands (H), Foods (F), Objects (O), and Animals (A). (b) In the initial training phase, participants were presented with trial-unique images from three categories of images. These contexts could be grouped together through an abstract LS representation based on the similarity of their category-value associations. (c) Participants were later trained on three new categories in a subset of the previous contexts. Grayed out columns indicate contexts that were left out of this phase. Thus, the values of new categories were trained in only one context in each LS cluster. (d) In the generalization phase, participants were asked to make decisions about the left out, novel context-category combinations without feedback. Participants had to use their knowledge of the LSs linking contexts together learned in the initial training. See Figure 1—figure supplement 1 for a table laying out the full experimental design over three experimental sessions.
Figure 1—figure supplement 1
Schematic showing design of whole experiment over multiple sessions.
Note that contexts seen in new category training and held out of generalization were not necessarily from Initial Training 3, but these are shown only as an example here. The mixed generalization phase did not take place in session 1. Only participants who passed the accuracy criterion in the generalization phase in session 1 completed the mixed generalization phase in the scanner, while participants who failed to meet this criterion completed this phase behaviorally. All other sections of the experiment took place outside of the scanner.
Figure 2 with 2 supplements
Learning curves and repetition effects from training and generalization phases for fMRI participants in session 2.
(a) Mean accuracy across participants for first six trials of the first nine mini-blocks of initial and new category training, as well as blocked generalization phase. Dashed line indicates chance-level performance. (b) Estimated exponential rate of change in mean accuracy in the first six trials for all mini-blocks in each phase. (c) Mean reaction times (RT) for each trial in each of the first nine mini-blocks for each of these phases. Shaded area represents standard error of the mean (SEM). Mean and standard error were calculated from the log-transformed RTs. (d) Estimated exponential rate of change in RT for first six trials within each phase for the first mini-blocks within the first three contexts in presentation order. For panels (b and d) each dot represents a single participant, horizontal and vertical bars represent the mean and SEM, respectively. *p<0.05, **p<0.01, Wilcoxon signed rank test, corrected for multiple comparisons. (e) RTs for trials from the pseudo-randomized generalization phase from sessions 2 and 3 during fMRI scanning where latent states remained the same or switched from the previous trial, while the context switched or stayed the same. RTs have been log-transformed and z-scored within session for each participant. Each line represents a participant, black dashed line indicates mean, and error bars indicate SEM. †p<0.05, †††p<0.0001, repeated-measures t-test. See Figure 2—figure supplement 1 for the mean accuracies for the fMRI and behavioral groups for each phase and session of the experiment. See Figure 2—figure supplement 2 for data from behavioral group participants who passed the generalization accuracy criterion in session 2.
Figure 2—figure supplement 1
Accuracy in all sessions of the experiment for behavioral and fMRI groups.
No statistical comparison between groups was made in the mini-blocked generalization phase of session 1 as accuracy in this phase was used to define these groups. Each point represents a single participant, horizontal bars represent the group mean and vertical bars indicate the standard error of the mean. **p<0.01, *p<0.05, ^p<0.1, Wilcoxon rank sum tests between groups.
Figure 2—figure supplement 2
Learning curves and repetition effects from training and generalization phases for behavioral participants who passed the accuracy criterion for generalization in session 2 (N = 9).
(a) Mean accuracy across participants for first six trials of the first nine mini-blocks of initial and new category training, as well as blocked generalization phase. Dashed line indicates chance-level performance. (b) Estimated exponential rate of change in mean accuracy in the first six trials for all mini-blocks in each phase, across participants. (c) Mean reaction times (RT) for each trial in each of the first nine mini-blocks for each of these phases. Shaded area represents standard error of the mean (SEM). Mean and standard error were calculated from the log-transformed RTs. (d) Estimated exponential rate of change in RT for first six trials within each phase for the first mini-blocks within the first three contexts in presentation order. For panels (b and d), each dot represents a single participant, and horizontal and vertical bars represent the mean and SEM, respectively. *p<0.05, Wilcoxon signed rank test, corrected for multiple comparisons, ^p<0.05, uncorrected. (e) RTs for trials from the pseudo-randomized generalization phase from sessions 2 and 3 where latent states remained the same or switched from the previous trial, while the context switched or stayed the same. RTs have been log-transformed and z-scored within session for each participant. Each line represents a participant, black dashed line indicates mean, and error bars indicate SEM. †p<0.05, ††p<0.01, repeated-measures t-test.
Figure 3 with 1 supplement
Reaction times (RT) for fMRI participants in mini-blocks during generalization in session 2, organized by the presentation of latent states (LS 1, 2, 3) and contexts (first and second).
Participants were significantly slower on the first trial of mini-blocks when encountering the first context in a LS compared to the second context in the same LS, but not in subsequent trials. (a) RT curves for mini-blocks organized by the order of LSs and contexts encountered. Lines indicate means, shaded area indicates the SEM. (b) Rates of exponential functions fit to these RT curves. Each dot represents a single participant. ††p<0.001, Wilcoxon signed-rank test, *p<0.05, Bonferroni-corrected for multiple comparisons, ^p<0.05, uncorrected. See Figure 3—figure supplement 1 for data from the behavioral group participants who passed the generalization criterion in session 2.
Figure 3—figure supplement 1
Reaction times (RT) for participants in the behavioral group who passed the generalization criterion in session 2 (N = 9) during mini-blocks during generalization in session 2, organized by the presentation of latent states (LS 1, 2, 3) and contexts (first and second).
Participants were significantly slower on the first trial of mini-blocks when encountering the first context in a LS compared to the second context in the same LS, but not in subsequent trials. (a) RT curves for mini-blocks organized by the order of LSs and contexts encountered. Lines indicate means, shaded area indicates the SEM. (b) Rates of exponential functions fit to these RT curves. Each dot represents a participant. Horizontal bars indicate the mean and vertical bars the SEM.
Figure 4 with 1 supplement
Computational modeling of reaction times (RT) during generalization phase for fMRI participants in session 2.
Each panel shows simplified schematics for spreading activation in four alternative model networks for a trial involving a decision about an image of an object in context A1 (i.e. A1-O-). (a) Conjunctive associative retrieval (CAR) model where each node represents a context-category conjunction with a particular value. (b) Independent associative retrieval (IAR) model where contexts and category-values are represented separately. (c) Latent state (LS) model where category-values are linked to a LS node, without any representation of context. (d) Hierarchical LS model with contexts clustered around LSs, and category-values linked to each context. Border and arrow colors signify mode of retrieval. Below each schematic are scatterplots for z-scored reaction times for all participants in all correct trials of the generalization phase and simulated reaction times from each model, color-coded by trial type. See Figure 4—figure supplement 1 for further details on fit of each model to the data. H, Hands; A, Animals; F, Faces; O, Objects. Values: +, positive expected value; −, negative expected value.
Figure 4—figure supplement 1
Supplementary analyses for computational modeling of reaction times (RTs).
Data are from session 2 of the mini-blocked generalization phase for participants in the fMRI group. (a) Root-mean squared deviation between data simulated from each model and participants’ z-scored reaction times for different trial-types. †p<0.05, ††p<0.01, repeated-measures t-tests, Bonferroni corrected for multiple comparisons. (b) Negative log-likelihoods for each participant in the fMRI group for conjunctive associative retrieval (CAR), independent associative retrieval (IAR), latent state (LS), and hierarchical latent state (HLS) models. Lower values indicate a better fit. (c) Values for three rate parameters governing rate of activation change in nodes due to direct (α1), mediated (α2), and incidental retrieval (α3). (d) Cross model comparison using best fitting parameters from 16 participants to generate simulated data, tested with alternative. Values indicate negative log-likelihood of model fits. **p=0.001, *p<0.05, Wilcoxon signed rank test, Bonferroni corrected for multiple comparisons.
Figure 5 with 5 supplements
Whole-brain representational similarity searchlight analysis for main effects of interest.
Each upper panel shows hypothesis representational dissimilarity matrix (RDM) for task factors. Lower panels show t-statistic map from a searchlight analysis testing these predictions in pattern activity projected onto inflated cortical surfaces. All maps are defined with a cluster forming threshold of p<0.001 and corrected for multiple comparisons with permutation tests for defining a cluster extent threshold at p<0.05. These maps include representations of (a) latent states (LSs), (b) contexts, and (c) value. Cluster extent threshold for each contrast is given by the value of k. A, B, C refer to distinct LSs. A1, B1, C1 and A2, B2, C2 refer to distinct contexts that belong to each of those LSs. F, Faces; A, Animals; O, Objects. +, positive value; −, negative value. See Figure 5—figure supplement 1 for comparison of LS map to 17 network parcellation by Yeo et al., 2011, Figure 5—figure supplement 2 for the statistical map for the item category regressor, and Figure 5—figure supplement 4 for interaction terms. See Figure 5—figure supplement 3 for comparison of main effects of interest within each session, and Figure 5—figure supplement 5 for a view of all regressors included in the multiple linear regression model.
Figure 5—figure supplement 1
Percentage overlap of latent state (LS) statistical map from Figure 5a with 17 network parcellation of functional connectivity networks from Yeo et al., 2011 based on the data of 1000 participants.
Numbers correspond to network identifications from the same study with percentage of voxels within each network shown in parentheses and represented by the pie chart. Surface renderings of the three networks that included the largest proportion of voxels are shown on an inflated cortical surface. Bottom right-hand corner shows the t-statistic map for the LS regressor from the representational similarity analysis searchlight for comparison, also shown in Figure 5a.
Figure 5—figure supplement 2
Whole brain searchlight analysis for image categories.
Left panel shows hypothesis representational dissimilarity matrix for categories (based on animacy). Right panel shows the cluster corrected t-statistic map from a searchlight analysis projected onto an inflated cortical surface. This statistical map was thresholded with a cluster forming threshold of p<0.001 and corrected for multiple comparisons using cluster-based permutation tests at p<0.05. The cluster extent threshold from this test is given by value of k.
Figure 5—figure supplement 3
Statistical maps for main effects of multiple regression model from searchlight representational similarity analyses (RSAs) estimated separately for data from sessions 2 and 3 and projected onto inflated cortical surfaces.
All statistical maps were thresholded with a cluster forming threshold of p<0.001 and corrected for multiple comparisons using permutation tests to find a cluster extent threshold (k) at p<0.05, except for the context effects in session 3 where no clusters passed this threshold (these data are shown at p<0.001, uncorrected).
Figure 5—figure supplement 4
Whole-brain searchlight analyses for interaction terms.
(a) Results showing where pattern activity was modulated by the negative interaction of latent state and value. Nuisance regressors for interactions between contexts and (b) value and (c) categories. All maps are defined with a cluster forming threshold of p<0.001 and corrected for multiple comparisons with permutation tests for defining a cluster extent threshold at p<0.05.
Figure 5—figure supplement 5
Hypothesis representational dissimilarity matrices (RDMs).
(a) All hypothesis RDMs included in the multiple regression analysis for searchlight and ROI-based representational similarity analyses (RSA). Color bar indicates dissimilarity. A, B, C refer to distinct latent states (LSs). A1, B1, C1 and A2, B2, C2 refer to distinct contexts that belong to each of those LSs. F, Faces; A, Animals; O, Objects. +, positive value; −, negative value. (b) Variance shared (R2) between hypothesis RDMs.
Figure 6 with 1 supplement
Representational similarity analysis results from all voxels included in regions of interest (ROIs).
Plots show distribution of beta coefficients across participants from multiple linear regression analyses comparing hypothesis representational dissimilarity matrices (RDMs) with empirical RDMs estimated for each ROI. Each point represents a single participant, with means represented by horizontal bars and SEM as vertical bars. (a) Orbitofrontal cortex (OFC), (b) ventral temporal cortex (VTC), and (c) hippocampus (HPC). *p<0.05, **p<0.01, ***p<0.0001 one-sample t-tests against zero. See Figure 6—figure supplement 1 for results of second-level tests on value and latent state terms restricted to the OFC ROI.
Figure 6—figure supplement 1
Results of representational similarity analysis searchlight results with explicit mask in orbitofrontal cortex.
Maps of t-statistics have been projected onto an inflated cortical surface to show effects for (a) latent states and (b) value. Both maps were defined with a cluster forming threshold of p<0.001 and corrected for multiple comparisons with permutation tests for defining a cluster extent threshold at p<0.05. The cluster extent threshold is given by the value of k.
Figure 7
Univariate whole-brain contrast of correct and erroneous responses projected onto inflated cortical surfaces and on a single axial slice showing orbitofrontal cortex and hippocampus.
This statistical map was defined with a cluster forming threshold of p<0.001 and corrected for multiple comparisons with permutation tests for defining a cluster extent threshold at p<0.05. Cluster extent threshold for each contrast is given by the value of k.
Author response image 1
Author response image 2
Author response image 3
Tables
Table 1
Sum of negative log-likelihood for four alternative models across participants and fraction of participants where each model was lowest in this measure in parentheses, for each group and session.
| Conjunctive associative retrieval | Independent associative retrieval | Latent state | Hierarchical latent state | |
|---|---|---|---|---|
| fMRI group – session 1 | 3537.78 (1/16) | 3227.38 (2/16) | 3428.23 (0/16) | 3061.04 (13/16) |
| fMRI group – session 2 | 3849.60 (0/16) | 3413.20 (4/16) | 3569.10 (0/16) | 3264.13 (12/16) |
| Behavioral group – session 2 | 2063.65 (0/9) | 1851.76 (0/9) | 1921.61 (0/9) | 1713.73 (9/9) |
-
Table 1—source data 1
Sum of negative log-likelihoods for individual participants in fMRI and behavioral groups for each model in each session.
- https://cdn.elifesciences.org/articles/63226/elife-63226-table1-data1-v2.zip
Additional files
-
Supplementary file 1
Activations passing permutation-based cluster correction for whole-brain representational similarity analysis.
All reported clusters were significant at the p<0.05, corrected for multiple comparisons after peak thresholding at p<0.001 and permutation-based cluster correction. The critical cluster extent threshold for each contrast is given by the value of k.
- https://cdn.elifesciences.org/articles/63226/elife-63226-supp1-v2.docx
-
Supplementary file 2
Activations passing permutation-based cluster correction for representational similarity analysis constrained to orbitofrontal cortex region of interest.
All reported clusters were significant at the p<0.05, corrected for multiple comparisons after peak thresholding at p<0.001 and permutation-based cluster correction within an explicit mask defining orbitofrontal cortex. The critical cluster extent threshold for each contrast is given by the value of k.
- https://cdn.elifesciences.org/articles/63226/elife-63226-supp2-v2.docx
-
Supplementary file 3
Activations passing permutation-based cluster correction for univariate contrast of correct and erroneous responses.
All reported clusters were significant at the p<0.05, corrected for multiple comparisons after peak thresholding at p<0.001 and permutation-based cluster correction. The critical cluster extent threshold for each contrast is given by the value of k.
- https://cdn.elifesciences.org/articles/63226/elife-63226-supp3-v2.docx
-
Transparent reporting form
- https://cdn.elifesciences.org/articles/63226/elife-63226-transrepform-v2.docx
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Neural representation of abstract task structure during generalization
eLife 10:e63226.
https://doi.org/10.7554/eLife.63226
