The topology of E/I recurrent networks regulates the effects of synaptic plasticity

Abstract

Synaptic plasticity is the primary mechanism for learning and memory in the brain. In recurrent neural networks, synaptic plasticity creates intricate feedback loops between population activity and connection strength. However, how the various topological features of brain networks, such as the diverse connectivity patterns of different neuron types, affect synaptic plasticity remains largely unknown. Here we investigate this question on the example of emergent excitatory and inhibitory co-tuning. This dynamical feature has been observed in cortical networks and was shown to be essential for efficient information processing. Computational models demonstrated that E/I co-tuning could arise from synaptic re-organization by a well-orchestrated plasticity protocol in low-noise feedforward networks. However, we show that the same plasticity protocol cannot give rise to E/I co-tuning in the presence of strong noise and unstructured recurrent connectivity. Using analytical methods and approximate Bayesian inference, we demonstrate that forming assembly structures in the recurrent connectivity can restore the ability of synaptic plasticity to produce E/I co-tuning, and we identify the optimal patterns for such co-tuning to emerge. In particular, we find that enhanced excitatory connectivity between similarly tuned neurons, combined with more homogeneous inhibitory connectivity, improves the ability of plasticity to produce co-tuning in an upstream population. Our results demonstrate how structured recurrent connectivity could control the ability of synaptic plasticity to adjust networks for efficient information processing.