Neurons and Cognition (q-bio.NC)

Abstract: Cognitive problem-solving benefits from cognitive maps aiding navigation and planning. Previous studies revealed that cognitive maps for physical space navigation involve hippocampal (HC) allocentric codes, while cognitive maps for abstract task space engage medial prefrontal cortex (mPFC) task-specific codes. Solving challenging cognitive tasks requires integrating these two types of maps. This is exemplified by spatial alternation tasks in multi-corridor settings, where animals like rodents are rewarded upon executing an alternation pattern in maze corridors. Existing studies demonstrated the HC - mPFC circuit's engagement in spatial alternation tasks and that its disruption impairs task performance. Yet, a comprehensive theory explaining how this circuit integrates task-related and spatial information is lacking. We advance a novel hierarchical active inference model clarifying how the HC - mPFC circuit enables the resolution of spatial alternation tasks, by merging physical and task-space cognitive maps. Through a series of simulations, we demonstrate that the model's dual layers acquire effective cognitive maps for navigation within physical (HC map) and task (mPFC map) spaces, using a biologically-inspired approach: a clone-structured cognitive graph. The model solves spatial alternation tasks through reciprocal interactions between the two layers. Importantly, disrupting inter-layer communication impairs difficult decisions, consistent with empirical findings. The same model showcases the ability to switch between multiple alternation rules. However, inhibiting message transmission between the two layers results in perseverative behavior, consistent with empirical findings. In summary, our model provides a mechanistic account of how the HC - mPFC circuit supports spatial alternation tasks and how its disruption impairs task performance.

Abstract: Learning relies on coordinated synaptic changes in recurrently connected populations of neurons. Therefore, understanding the collective evolution of synaptic connectivity over learning is a key challenge in neuroscience and machine learning. In particular, recent work has shown that the weight matrices of task-trained RNNs are typically low rank, but how this low rank structure unfolds over learning is unknown. To address this, we investigate the rank of the 3-tensor formed by the weight matrices throughout learning. By fitting RNNs of varying rank to large-scale neural recordings during a motor learning task, we find that the inferred weights are low-tensor-rank and therefore evolve over a fixed low-dimensional subspace throughout the entire course of learning. We next validate the observation of low-tensor-rank learning on an RNN trained to solve the same task by performing a low-tensor-rank decomposition directly on the ground truth weights, and by showing that the method we applied to the data faithfully recovers this low rank structure. Finally, we present a set of mathematical results bounding the matrix and tensor ranks of gradient descent learning dynamics which show that low-tensor-rank weights emerge naturally in RNNs trained to solve low-dimensional tasks. Taken together, our findings provide novel constraints on the evolution of population connectivity over learning in both biological and artificial neural networks, and enable reverse engineering of learning-induced changes in recurrent network dynamics from large-scale neural recordings.