V-121 | Network connectivity and dimensionality in heterogeneous neural networks

Theoretical and Computational Neuroscience
Author: Martina Acevedo | Email: martina.acevedo@ib.edu.ar

Martina Acevedo^1°, Soledad Gonzalo Cogno^2°,Germán Mato^1°3°4°

^1° Instituto Balseiro, Universidad Nacional de Cuyo and CNEA, Argentina.
^2° Kavli Institute for Systems Neuroscience and Centre for Algorithms in the Cortex, Norwegian University of Science and Technology (NTNU), Trondheim, Norway.
^3° Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Argentina.
^4° Department of Medical Physics, Centro Atómico Bariloche, San Carlos de Bariloche, Argentina.

Brain function emerges from the coordinated activity of interconnected neurons. Neuronal coordination gives rise to patterns of population activity that span different degrees of dimensionality. While neurons exhibit a wide range of intrinsic firing properties, theoretical studies have typically investigated the link between network connectivity and dimensionality using the same single-cell dynamics. Hence, the effect of heterogeneity in single-cell dynamics on this link remains unknown. To address this question, we developed recurrent neural network (RNN) models composed of leaky integrate-and-fire (LIF) neurons, quadratic integrate-and-fire (QIF) neurons, or a mix of both.
We found that when the RNN was in a chaotic state, the dimensionality of the activity depended on the RNN composition and recurrency. As weights increased, the dimensionality always gradually decreased, but with QIF networks keeping a higher dimensionality than LIF ones. We next used FORCE learning to enforce oscillatory activity in the RNN units. After learning, the variability in the weights was smallest in LIF networks, largest in QIF networks, and intermediate in mixed networks. The organization of the weights into connectivity motifs also varied with the network composition, with bidirectional connections being more abundant in networks with more QIF neurons. All in all, our results suggest that the presence of heterogeneous single cell dynamics shapes both neural connectivity and network dynamics.