Dynamic updating of hippocampal object representations reflects new conceptual knowledge

Abstract

Concepts organize the relationship among individual stimuli or events by highlighting shared features. Often, new goals require updating conceptual knowledge to reflect relationships based on different goal-relevant features. Here, our aim is to determine how hippocampal (HPC) object representations are organized and updated to reflect changing conceptual knowledge. Participants learned two classification tasks in which successful learning required attention to different stimulus features, thus providing a means to index how representations of individual stimuli are reorganized according to changing task goals. We used a computational learning model to capture how people attended to goal-relevant features and organized object representations based on those features during learning. Using representational similarity analyses of functional magnetic resonance imaging data, we demonstrate that neural representations in left anterior HPC correspond with model predictions of concept organization. Moreover, we show that during early learning, when concept updating is most consequential, HPC is functionally coupled with prefrontal regions. Based on these findings, we propose that when task goals change, object representations in HPC can be organized in new ways, resulting in updated concepts that highlight the features most critical to the new goal.

Keywords: attention; category learning; computational modeling; fMRI; hippocampus.

Fig. 1.

Experiment schematic and behavioral performance. (A) Participants learned to classify eight insect images according to two rules through feedback-based learning. On every trial, an insect image was presented (3.5 s) and participants made classification responses according to the current task. After a delay (0.5–4.5 s), feedback consisting of the insect image, accuracy, and the correct response was shown (2 s). The next trial began after a variable delay (4–8 s). For both tasks, participants responded to all eight stimuli over 16 repetitions. (B) The stimuli consisted of insects with three binary features (thick/thin legs, thick/thin antennae, and pincer/shovel mouths). The stimulus set consisted of eight images representing all combinations of the three binary features. The two classification tasks required attention to different features: the type 1 problem was based on one feature (e.g., the antennae) and the type 2 problem was an exclusive disjunction classification based on a combination of two features (e.g., the mouth and legs). The feature-to-task mappings and order of the learning tasks were counterbalanced across participants. (C) The average probability of a correct response across the 16 learning repetitions is plotted for both tasks. Error bars represent 95% CIs around the inflection point of the bounded logistic learning curves. The shaded ribbons represent 95% CIs of the mean.

Fig. 2.

Schematic of learning model and model predictions. (A) The learning model consists of three main components (see SI Experimental Procedures, Computational Learning Model for model formalism). First, the sensory input of the three features is attenuated by receptive fields tuned according to attention weights (λ_i). The attention component alters the perceptual representation of the stimulus toward task-diagnostic information. Second, stored knowledge represented by clusters of weighted features compete to be activated by the attention-biased input. The cluster most similar to the attention-biased input wins and activates the class unit. Third, the activated class unit serves as input to a decision component that generates a response. Trial-to-trial, the model learns through feedback by updating the attention weights and the weights connecting clusters to the class unit and whether an existing cluster is updated or a new cluster is recruited. (B) The model was fit to participants’ learning performance (Fig. 1C) and the final attention weights (λ_i) for each dimension were extracted for both tasks. The relative attention weights for each task are depicted in the radar plots (dotted lines show participant weights, bold lines show group means). (C) Matrices depict the average model predictions for the pairwise similarities between the stimuli for the two tasks. Task-specific similarity predictions for each participant were generated by extracting cluster activations for each stimulus at the end of learning. Pearson correlations were then calculated for each stimulus pair, and averaged across participants. The similarity matrices characterize the task-specific conceptual representations underlying classification decisions. Stimuli in the same class for a given task are marked by text color on the matrix axes.

Fig. 3.

Schematic of model-based RSA. Model predictions and neural measures of stimulus similarity were extracted from the second half of both tasks. For each participant, the learning model was fit to behavior and used to generate representational similarity spaces (Fig. 2C). A searchlight method was used to generate corresponding neural similarity matrices within the hippocampus (highlighted in green) by correlating voxel activation patterns within each searchlight sphere (3-voxel radius) for all stimulus pairs from fMRI data recorded during the latter half of the task. The correspondence between model and neural similarity matrices across both tasks was assessed with Spearman correlation.

Fig. 4.

Model-based RSA and learning-related connectivity results. (A) Neural representations in left anterior HPC were consistent with model predictions of attention-weighted conceptual coding (peak x = −25, y = −15, z = 17, 161-voxel cluster extent; voxelwise thresholded at P < 0.005 and small volume corrected at P < 0.05 for HPC). (B) Stimulus-specific neural representations from the HPC region in A were used to estimate attention weights to the three feature dimensions. These neurally derived attention weights were then projected into feature space to demonstrate the attentional tuning across tasks. Each point represents a stimulus and is colored according to the class membership for the task. The attention-weighted spaces are a visual depiction of the model-based RSA results (i.e., they are not an independent analysis) and show how attention is tuned across tasks to reconfigure stimulus space into task-relevant conceptual space. (C) Regions in PFC and occipital cortex showed significantly greater functional coupling with the HPC region identified by model-based RSA during early versus late learning (voxelwise thresholded at P < 0.005, whole brain cluster extent corrected at P < 0.05).