Representations of Temporal Community Structure in Hippocampus and Precuneus Predict Inductive Reasoning Decisions

Abstract

Our understanding of the world is shaped by inferences about underlying structure. For example, at the gym, you might notice that the same people tend to arrive around the same time and infer that they are friends that work out together. Consistent with this idea, after participants are presented with a temporal sequence of objects that follows an underlying community structure, they are biased to infer that objects from the same community share the same properties. Here, we used fMRI to measure neural representations of objects after temporal community structure learning and examine how these representations support inference about object relationships. We found that community structure learning affected inferred object similarity: When asked to spatially group items based on their experience, participants tended to group together objects from the same community. Neural representations in perirhinal cortex predicted individual differences in object grouping, suggesting that high-level object representations are affected by temporal community learning. Furthermore, participants were biased to infer that objects from the same community would share the same properties. Using computational modeling of temporal learning and inference decisions, we found that inductive reasoning is influenced by both detailed knowledge of temporal statistics and abstract knowledge of the temporal communities. The fidelity of temporal community representations in hippocampus and precuneus predicted the degree to which temporal community membership biased reasoning decisions. Our results suggest that temporal knowledge is represented at multiple levels of abstraction, and that perirhinal cortex, hippocampus, and precuneus may support inference based on this knowledge.

Figure 1.

(A) Temporal community structure with novel 3-D objects assigned to its 21 nodes. The connections between nodes indicate which objects may follow one another in succession in a structured sequence. Nodes are divided into three distinct communities (purple, red, and green). Darker circles indicate central nodes, that is, objects that are connected to all other objects in the same community. Lighter circles indicate boundary nodes, that is, objects that are connected to the five central nodes and another boundary object in a different community. The arrows show an example sequence through the network, starting with the outlined object; arrow color corresponds to position within the sequence. (B) A structured sequence of objects corresponding to the path illustrated in (A). For illustrative purposes, background color corresponds to the node colors in (A). In the actual experiment, objects were presented on a white background.

Figure 2.

(A) Participants first viewed structured sequences of objects derived from the temporal community structure graph. Next, participants were scanned during presentations of additional structured sequence blocks, which were intermixed with scrambled sequence blocks in which objects were presented in a random order that did not follow the temporal community structure. (B) Outside of the scanner, participants performed an inference task to measure how temporal structure learning affected reasoning about object properties. Participants were told that a cue object was found in one of three environments (ocean, desert, or forest) and asked to choose which of two objects (shown below the cued object) could also be found there. For each trial, one of the object choices shared the same temporal community as the cue, and the other object choice did not. There were three types of trials (central, boundary 1-away, and boundary 2-away), which varied in the degree to which the choice objects were directly or indirectly connected with the cue object in the structure graph. (C) The parsing task measured subjective perception of event boundaries. Participants were presented with object sequences that were generated either through a structured walk through the graph or a Hamiltonian walk in which the nodes were visited one by one in turn. Participants were told to press a button (to indicate an event boundary) whenever they felt a subjective shift in the sequence. In the example sequence, the object’s background color (not shown to participants) indicates its community; dark squares indicate central nodes, and light squares indicate boundary nodes. (D) The grouping task measured the learned similarity of the objects and examined whether temporal knowledge generalized to a spatial grouping task. Participants were shown a grid with the 21 objects they had seen in the task randomly placed on it and were asked to group the objects based on their experience.

Figure 3.

SRs after learning given different values of the $\text{α}$ and $\text{γ}$ parameters, after simulation of the object sequences viewed by an example participant. Within each SR, each row shows the expected count of future object presentations, discounted based on the discounting factor $\text{γ}$, for a given starting object. Circles indicate the object community (red, purple, green) and whether each object is a central node (dark) or boundary node (light). For display purposes, each matrix was divided by the maximum value in that matrix. We found that the clearest learning of the community structure appeared for relatively low values of $\text{α}$ and relatively high values of $\text{γ}$.

Figure 4.

Behavioral measures of community structure learning. (A) Parsing probability during sequence viewing. Sequences were either Hamiltonian or random walks through the temporal community structure graph. Parsing occurred more frequently after transitions between communities compared with other transitions. Points indicate individual participants; error bars indicate bootstrap 95% confidence intervals. (B) In the grouping task, participants placed within-community objects nearer to one another than across-community objects. (C) Participants showed evidence of a bias toward inferring the same properties for objects from the same communities. Positive scores indicate bias toward inferring common properties of objects from the same community. (D) Parsing performance (i.e., event segmentation after community transitions relative to segmentation after other transitions) predicted smaller distances between objects in the same community on the grouping task. Both measures were residualized against rotation detection performance. Shaded area indicates 95% confidence intervals for the best-fitting linear regression. (E) Parsing performance predicted temporal bias on the inference task. Both measures were residualized against rotation detection performance. *$p<.05$.

Figure 5.

Models of inductive inference task performance based on learned similarity of object pairs. (A–B) SR matrix and model fit to inference behavior. (A) SR matrix after learning for one example participant (the same participant as in Figure 3) based on the object sequences that they observed, with a learning rate of 0.1 and best-fitting discounting factor of 1.0. The matrix shows the expected temporally discounted frequency of visiting each inference object after viewing the cue object. Circles indicate the object community (red, purple, green) and whether each object is a central node (dark) or boundary node (light). (B) Fit of the SR model to average temporal bias for each inference test type. Bars show observed performance (with bootstrap 95% confidence intervals); red dots show model estimates. The model incorrectly predicts that temporal bias should be low on the boundary 2-away trials. (C–D) Within-community similarity model. (C) Model association matrix. Objects in the same community have a similarity of one, and objects in different communities have a similarity of zero. (D) The model incorrectly predicts that there will be equal bias for each trial type. (E–G) Hybrid SR-community model. The hybrid model simulates inference as being influenced by a weighted combination of SR and within-community similarity. (E) SR matrix for the same example participant as in (A), with the best-fitting discounting factor of 0.974. (F) The hybrid model correctly fits the differences in temporal bias between the trial types. (G) Based on the model fit, the central and boundary 1-away inference trials weight the SR more heavily, whereas the boundary 2-away trials weight only the within-community similarity matrix. C = central; B1 = boundary 1-away; B2 = boundary 2-away.

Figure 6.

Right anterior hippocampus similarity is greater for objects in the same community compared with objects in different communities.

Figure 7.

(A) A greater difference between within- and across-community similarity in left perirhinal cortex predicted smaller distance between same-community objects in the grouping task. (B) Left precuneus showed a greater difference between within- and across-community similarity for participants that were more biased by community structure on the inference task.

Figure 8.

(A) Community structure in the left precuneus cluster was uniquely related to temporal bias on the boundary 1-away inference trials. (B) Community structure in the right anterior hippocampus cluster was uniquely related to temporal bias on the boundary 2-away inference trials.