Abstract
The firing rate of macaque lateral intraparietal (LIP) neurons encodes the decision variable (DV) for perceptual decisions reported through saccadic eye movements in a random dot direction discrimination task (Shadlen & Newsome, 2001). In circuit models for these decisions, neural ensembles that encode actions compete to form decisions. Consequently, DVs are represented as partially enabled action plans, where ensembles increase their average responses for stronger evidence supporting their preferred actions. As another consequence, DV representation and readout are implemented similarly for different inputs and task contexts when decisions are communicated through the same actions. Here, we examined these assumptions by comparing LIP responses between the motion discrimination task and a novel face discrimination task. Unlike in the motion task, average LIP firing rates during face discrimination were lower for stronger stimuli supporting the preferred saccade target of the recorded neurons, contradicting existing theories. This marked difference in average responses, however, did not indicate different underlying computations between the tasks; population response patterns in both tasks monotonically encoded the DV on a curved manifold in the state space, representing the integration of sensory evidence. This curved manifold rotated and shifted in a task-dependent manner, leading to the opposing trends of average firing rates in the two tasks. These newly discovered properties of LIP activity are not explained by existing circuit models, necessitating development of new models that incorporate task-dependent computations. Furthermore, the curvature of manifold encoding the DV was not limited to LIP; similar curved manifolds were discovered in the lateral and medial frontal cortical regions. This indicates a ubiquitous computational mechanism across the frontoparietal regions that underlies the observed geometry of the DV encoding.