Context-Dependent Epoch Codes in Association Cortex Shape Neural Computations

Neural circuits adapt their computations to perform cognitive functions within and across tasks. These adaptations are vital for decision making, which often requires the integration of past information and present context to generate appropriate behaviors. Neurons in association cortices exhibit selective activity for task-relevant signals such as current and past choices and outcomes, yet how such selectivity is modulated by the task context is unclear. Interestingly, neuronal activity encodes time or epoch within the task as strongly as conventional task variables. Here we examined the intriguing hypothesis that such condition-independent "epoch codes" may play a role in shaping neural computations across time and contexts. We analyzed >800 single-neuron recordings from multiple association cortical areas — LIP, dlPFC, ACC, and SEF — while monkeys performed two different tasks: (i) a competitive
"matching pennies'' decision-making task that relies on the integration of choice and outcome memory across trials, and (ii) a control "search" task that does not require across-trial memory. We found that neuronal trace of choice and outcome is gated by an epoch code multiplicatively rather than additively. Furthermore, task-relevant memory signals are stronger and the epoch codes are modulated more during matching pennies than during search task. We also found significant
differences in the geometry of population epoch-code trajectories across cortical areas. Notably, epoch codes in LIP and ACC were lower-dimensional, indicating less heterogeneity in patterns across neurons than those in dLPFC and SEF. Epoch codes in LIP exhibited less modulation by task context relative to the other regions. By separating epoch code from task-variable coding, we tested how epoch codes, and their contextual modulation, relate to task-relevant computations. Similarly, this characterization of contextual epoch codes across association cortical areas also revealed
computational principles for dynamical network models of flexible cognitive processes.