Circuits in PFC are believed to integrate sensory evidence and carry out context-dependent computations. In a recent study by Mante et al.[1], monkeys were trained to perform a dual-context decision-making task, learning to select a contextually-relevant stimulus while ignoring an irrelevant one. A non-linear RNN model was proposed, to explain how the circuit could be capable of integrating the relevant input while dynamically washing away the irrelevant one. The mechanism could be understood in terms of selection plus integration of input vectors mediated by the left and right eigenvectors of the underlying linearised system. In the present study, we continue this work complementing it with a new analysis approach by inferring a linear dynamical system (LDS) model [2] directly from the data. The model-generated population trajectories capture the main features of the original trajectories. The dynamics suitably adapt to each context, changing the orientation of the left eigenvectors so that relevant inputs are more amplified. This amplification, however, is being carried along several slow decaying modes. Importantly, despite having fitted a separate model to each context, the inferred input directions are stable across contexts. Furthermore, we were able to find a correspondence between the task-relevant dimensions found by regression in the real data[1] and the learned model parameters, as regression stimulus vectors were mostly aligned with the corresponding model input vectors. Finally, we also analyse the impact of different observational noise models on the inferred underlying dynamical system. The particular dynamical solution that this model finds provides crucial support for the notion that changes in integration, rather than rebalanced input strengths, underlies the flexible behaviour. Specifically, it appears consistent with contextually-dependent alignment of inputs onto directions of dynamical persistence.

[1] Mante, V., Sussillo, D., Shenoy, K. V., and Newsome, W. T. , Nature 503(7474):78-84 (2013).
[2] Macke, J. H., Busing, L., Cunningham, J. P., Yu, B. M., Shenoy, K. V., Sahani, M. , NIPS (2011).