Functional populations in prefrontal cortex related to working memory encoding and maintenance

Abstract

Nonlinear mixed selectivity, with neurons responding to diverse combinations of task-relevant variables, has been proposed as a key mechanism to enable flexible behavior and cognition. However, it is debated whether the structure of neural population responses in fronto-parietal cortices is better described as random mixed-selective or as non-random, that is, in terms of multiple subpopulations with stereotypical response profiles. Here, we show that neural activity in the macaque prefrontal cortex during a working memory and a visual response task is organized into subpopulations that provide a comprehensive description of the low-dimensional population dynamics. First, analysis of the demixed Principal Components shows that the neural code faithfully represents stimulus identity, task condition, and elapsed time during the trial. Second, a model-free analysis of the population structure reveals a significant degree of clustering, implying a non-random distribution of feature selectivity that is incompatible with random mixed selectivity. Closer inspection shows that stimulus-selective neurons also tend to be task-selective. Third, examining the contribution of stimulus-selective neurons to task-condition-related variance reveals two contrasting activity profiles that correspond to functionally different populations. One population responds during visual stimulation while the other activates during memory maintenance. Finally, the observed neural geometry explains how stable task and stimulus information can be read out from the population response using a linear decoder. Our results highlight that despite the heterogeneity of prefrontal responses during working memory, neurons do not represent random mixtures of task features but are structured according to neural subpopulations.