Map Making: Cognitive maps, cognitive control, and inference

Abstract

Recent findings suggest the hippocampal-entorhinal (HPC-ERC) system may serve a general mechanism for representing and navigating cognitive maps of non-spatial tasks. These map-like representations can be used to guide flexible goal-directed decisions. However, it is unclear whether this system, and interconnected medial prefrontal cortex (mPFC), use the same principles to guide everyday decisions in the absence of continuous sensory feedback – such as whom to collaborate with. In my talk I will present a series of studies designed to address these gaps. 

In study 1 (Park et al., Neuron, 2020) participants first separately learned the rank of neighboring people on each of two social dimensions–popularity and competence–in two separate groups. Participants could infer the overall structure of each group hierarchy, which was never shown, through transitive inferences. Next, they learned the relative rank of select individuals (“hubs”) in each group with the other group, creating a unique associative path between groups. Finally, they made inferences about the rank of novel pairs between the two groups. Pattern similarity analyses showed that people who were more proximal in Euclidian distance in the mentally reconstructed 2-D space were represented progressively more similarly in HPC, ERC, and orbitofrontal cortex, consistent with a cognitive map of the abstract 2-D social hierarchy. Notably, cognitive control demands produced warping in the representational geometry of HC such that distances between congruent pairs (those requiring context-invariant responses) were longer than equidistant incongruent pairs (requiring context-dependent responses). 

It could further be shown that two complementary learning strategies implemented in artificial recurrent neural networks (RNNs) trained to perform our task in a manner analogous to our human subjects can both achieve near-optimal performance in our task: 1) an explicit episodic retrieval of specific sets of past comparisons between pairs and 2) direct inferences over a 2D cognitive map formed through error-driven learning (back-propagation) in an RNN. Interestingly, retrieval of the hub in the hippocampus is predicted by the first system, while the second system recapitulated the 2D map-like representations of the social hierarchy in the embedding of the faces of the hierarchy in the hidden layer of the RNN, and also consistently showed warping along the context-invariant “congruent” axis (Russin et al., Cosyne, Proc. Cog. Sci. Soc., 2021). These results shed light on the mechanisms by which abstract and discretely sampled social cognitive maps are constructed, represented, and used for task-relevant inferences in the human brain. 

In study 2 (Park et al., Nature Neuroscience, 2021) we investigated the neural code that might underpin discrete decisions about individuals in the social network. In each trial, an entrepreneur was presented with two potential collaborators. Participants were asked to choose the better partner between two for a given entrepreneur by comparing the ‘growth potential (GP)’ of two pairs. First, we replicated the finding that the level of pattern dissimilarity in the HPC and ERC increased with the pairwise Euclidean distance between entrepreneurs in the 2-D social network, suggesting that separately learned dimensions are integrated into a 2-D cognitive map. Second, we found that the ERC, mPFC, temporoparietal junction area (TPJ), and superior temporal sulcus all display hexadirectional signals to encode the direction of inferred trajectories between entrepreneurs over the abstract social space. Third, we found the growth potential (equivalent to a pair’s value) was independently encoded in overlapping regions of mPFC and TPJ, and these effects were further modulated by the grid-like code, suggesting a dependence of growth potential coding on the grid-like coding. Finally, we found that the ERC and vmPFC encode the difference in GP between pairs at the time of decisions, suggesting their roles in comparing the values of each decision option from the multidimensional cognitive map. Our findings show that a grid-like code in the human brain is extended to encode inferred direct trajectories over an abstract and discrete social space during decision making, which may suggest a general mechanism for how the brain implements model-based decisions.