Dissecting modular recurrent neural networks trained to perform un-cued task switching

Animals can switch rapidly between multiple well-learned tasks without being explicitly instructed on which task to perform, a cognitive function termed un-cued task switching. This function relies on higher-order cortical areas to maintain and update the task rule representation, and use this representation to gate the sensorimotor pathway. How different brain regions and neuronal types interact to achieve this function is not well understood. We investigated a collection of modular recurrent neural networks (RNNs) trained to perform an analog of the Wisconsin Card Sorting Test (WCST), a classic task involving un-cued task switching. Each network consists of two interacting modules: a “prefrontal cortex (PFC)” module which encodes the task rule and a “sensorimotor” module which performs the sensorimotor mapping conditioned on the rule. Each module contains a canonical circuit with excitatory neurons and three types of inhibitory neurons. We extensively analyzed dozens of networks trained with different hyperparameters and discovered common underlying circuit motifs. Specifically, two types of neuronal responses emerge among the excitatory neurons in the PFC module - one that persistently encodes the rule and the other that transiently encodes the conjunction of negative feedback and rule. The PFC recurrent connectivity matrix can be equivalently described by a low-dimensional connectivity between six neuronal groups. In the sensorimotor module, a geometric relationship between input weight vectors from different stimulus features ensures appropriate action selection. The PFC module gates the sensorimotor module by sending long-range projections that connect excitatory neurons with similar rule preferences. In addition, excitatory neurons in the sensorimotor module are disinhibited by somatostatin (SST) neurons with the opposite rule preference. Lastly, the extent to which different dendritic branches encode different task rules is largely determined by the sparsity of the connections from the SST neurons to the excitatory neurons. Collectively, our study revealed distinct intra-areal and inter-areal neural circuit mechanisms underlying un-cued task switching.