Theoretical and Computational Neuroscience
Author: Federico Szmidt | Email: fszmidt@gmail.com
Federico Szmidt1°, Camilo J. Mininni1°
1° Instituto de Biología y Medicina Experimental
The hippocampal formation is crucial for the neural representation of space and time. Particularly in the hippocampus, CA1 and CA3 neurons use sparse coding to represent these variables, with place cells firing in specific spatial regions and time cells firing at specific times within an interval. However, many neurons display a mixed representation, firing in response to both a particular time and place. While several models have been proposed to explain the emergence and mechanisms of place and time cells separately, mixed representation remains unexplained.
In this work, we trained recurrent neural networks with two interconnected modules: one encoding an agent’s position by integrating speed from simulated trajectories and visual inputs, and another encoding elapsed time between discrete stimuli. The architecture promotes neural competition dynamics within each module. The trained networks exhibited sparse coding similar to that observed in hippocampal place cells and time cells, with many neurons demonstrating mixed selectivity. The networks also shared other properties with biological systems, such as place field size distributions and the occurrence of Weber’s Law in time cells. Finally, we explored how network connectivity and dynamics give rise to these properties. Our findings suggest that sparse coding and mixed representation for space and time naturally arise from a joint space-time optimization problem involving competition dynamics.