Neuronal implementation of the representational geometry in prefrontal working memory

Representational geometry is a prevalent approach to abstract and investigate population coding derived from the activities of individual neurons. However, the same representational geometry can be implemented by neurons in different ways, which implies different potential readout mechanisms. Leveraging the biological constraint of sparsity of readout synapses, we identified the biologically meaningful sparse components that express the geometry of working memory representations in primate prefrontal cortex (PFC). The dominant neuron groups corresponding to each component had distinct electrophysiological properties. We found that memorized information was represented in a sequential manner by these neuron groups that followed the task requirements. A recurrent neural network (RNN) model was trained to reproduce the firing rates in the PFC population. The RNN's accuracy dropped when the sequential sparse implementation in the data was destroyed, suggesting the necessity of this specific implementation to the observed activity in PFC. This study provides the perspective of neuronal implementation as an important complement to representational geometry, which helps bridge the gap between single-neuron activity and population-level dynamics.