Recurrent neural network models of multi-area computation underlying decision-making

Cognition emerges from coordinated computations across multiple brain areas. However, elucidating these computations within and across brain regions is challenging because intra- and inter-area connectivity are typically unknown. To study coordinated computation, we trained multi-area recurrent neural networks (RNNs) to discriminate the dominant color of a checker-board and output decision variables reflecting a direction decision, a task previously used to investigate decision-related dynamics in dorsal premotor cortex (PMd) of monkeys. We found that multi-area RNNs, trained with neurophysiological connectivity constraints and Dales law, recapitulated decision-related dynamics observed in PMd. The RNN solved this task by a dynamical mechanism where the direction decision was computed and outputted, via precisely oriented dynamics, on an axis that was nearly orthogonal to checkerboard color inputs. This orthogonal direction information was preferentially propagated through alignment with inter-area connections; in contrast, color information was filtered. These results suggest that cortex uses modular computation to generate minimal sufficient representations of task information. Finally, we used multi-area RNNs to produce experimentally testable hypotheses for computations that occur within and across multiple brain areas, enabling new insights into distributed computation in neural systems.