Conference on Cognitive Computational Neuroscience

Recent work has shown that abstract, non-spatial relationships between task-relevant states or entities are organized into map-like neural representations. Here, we investigate how these map-like representations interact with changing task goals in the context of cognitive control, where the features most relevant to the current goal benefit from top-down biasing. Classic computational neuroscience studies of cognitive control have focused on explicitly presented categorical features rather than map-like representations retrieved from memory, and have typically found facilitation of task-relevant features and suppression/compression of task-irrelevant features. Here, we explore the relationship between cognitive control and the geometry of map-like representations by combining neural network models and fMRI of the same task. Previously, we found that although only one of two task attributes was behaviorally relevant for current decisions, hippocampus (HC), entorhinal cortex (EC), and orbitofrontal cortex (OFC) spontaneously organized pairwise relationships into 2D map-like representations. Consistent with the predictions of the neural-network models, new analyses of the fMRI data show that task-irrelevant dimensions were compressed relative to task-relevant dimensions dynamically as a function of which dimension is currently relevant, in dorsomedial and dorsolateral frontal (DLFC and DMFC) and posterior and medial parietal cortex (PMC). Furthermore, the model’s underlying 2D representations were also affected by task demands in a different way: representations were warped along the 2D axis that remains unchanged across conditions requiring focus on each dimension separately. This finding was confirmed by fMRI analyses showing that this same warping phenomenon occurs in the HC, and that the degree of warping was correlated with individual differences in cognitive control. Further simulations showed that this warped geometry reflects the natural tendency of neural networks to learn context-invariant maps, consistent with behavioral and fMRI results. The stronger reactivation of context information in the HC to mitigate the conflicts was associated with the less the warping in the representations across individuals.