Linear dynamics of evidence integration in contextual decision making
Abstract: Individual neurons in Prefrontal Cortex (PFC) exhibit a vast complexity in their responses. Central in Neuroscience is to understand how their collective activity underlies powerful computations responsible for higher order cognitive processes. In a recent study (Mante et al., 2013) two monkeys were trained to perform a contextual decision-making task, which required to selectively integrate the relevant evidence -either the color or the motion coherence of a random dots stimulus- and disregard the irrelevant one. A non-linear RNN trained to solve the same task found a solution that accounted for the selective integration computation, which could be understood by linearizing the dynamics of the network in each context. In this study, we took a different approach by explicitly fitting a Linear Dynamical System (LDS) model to the data from each context. We also fitted a novel jointly-factored linear model (JF), equivalent to the LDS but with no dynamical constraints and able to capture arbitrary patterns in time. Both models performed analogously, indicating that PFC data display systematic dynamics consistent with the LDS prior. Motion and color input signals were inferred and spanned independent subspaces. The input subspaces largely overlapped across contexts along dimensions that captured coherence and coherence magnitude related variance. The dynamics changed in each context so that relevant stimuli were strongly amplified. In one of the monkeys, however, the integrated color signal emerged via direct input modulation. The integration took place within subspaces spanned by multiple slow modes. These strongly overlapped along a single dimension across contexts, which was consistent with a globally identified decision axis. Interestingly, irrelevant inputs were not dynamically discarded, but were also integrated, although in a much lower extent. Finally, the model reproduced the main dynamical features of the population trajectories and accurately captured individual PSTHs. Our study suggests that a whole space of sensory-related input signals invariantly modulates PFC responses and that decision signals emerge as the inputs are shaped by a changing circuit dynamics. Our findings imply a novel mechanism by which sensory-related information is selected and integrated for contextual computations.