Item

Abstract

The connectivity of cortical networks often includes functional clusters of strongly intercon- nected, similarly tuned neurons. While these assemblies are known to include both excitatory and inhibitory neurons, there is a high variability of specificity and connectivity patterns of different neuron types. Although there are theoretical studies linking the presence of excita- tory assemblies with the emergence of rich dynamics in spiking networks [1], the impact of neuron type-specific connectivity patterns on the dynamics and computational capabilities of recurrent networks remains poorly understood. Here, we use mean-field theory to assess the impact of E and I assemblies of varying strengths on the dynamics of a balanced recurrent network. We demonstrate that in networks with sufficiently strong coupling, different levels of clustering among E and I neurons can control the distance from a chaotic transition. In particular, we show that relatively weak I assemblies combined with stronger E assemblies can maintain the network’s dynamics at the edge of chaos. We then evaluate how the topology-induced dynamics impact the computational capabilities of the recurrent network using a reservoir computing model. Specifically, we train our network on the complex task of simultaneously predicting multiple chaotic timeseries. We find that the performance can be optimized for any given average coupling strength (within a reasonable range) by adjusting the relative strength of E and I assemblies. Finally, we use simulation-based inference [2] to evaluate the interaction between the speci- ficity of E and I input, the relative strength of E and I assemblies, and the network performance. We find that in near-chaotic dynamical regimes, broader E than I input can stabilize activity and boost performance. Our findings provide a description of how diverse connectivity between different neuron types can generate complex dynamics associated with superior performance in computational tasks. The connectivity patterns we identify largely coincide with experimentally observed structures in the mammalian cortex, suggesting a potential link between network topology, population dynamics, and the ability to perform challenging computations.