Emergent functional organisation and population geometry of recognition memory for numerosity in Recurrent Neural Networks

Recognition memory for numerical quantities requires neural systems to encode, maintain, and compare numerosity information over time. Neurophysiological studies in non-human primates have identified specialized neuronal populations associated with these processes. Here, we use recurrent neural networks (RNNs) as computational models to investigate whether and how similar functional organization and population-level representations emerge through task learning. Networks were trained on a delayed match-to-numerosity task and analyzed at both the single-unit and population levels. We characterized neuronal selectivity, delay-period activity, representational geometry, and functional organization, comparing the resulting network dynamics with published electrophysiological findings. Beyond static population geometry, we tracked the temporal evolution of low-dimensional neural manifolds throughout the task, relating changes in representational structure to the computational demands of memory encoding, maintenance, and comparison. The trained networks developed stable memory representations and distinct functional neuronal populations resembling experimentally observed response profiles. These findings suggest that task-driven optimization can give rise to biologically relevant functional organization and dynamic neural representations, providing a computational perspective on numerosity recognition memory and its underlying cortical population dynamics.