Emerging assembly structures in trained spiking neural networks

Authors: Julia Gygax, Friedemann Zenke

Presentation type: Flash talk at SNUFA 2025 online workshop (5-6 Nov 2025)

Abstract

How circuits of spiking neurons represent information in the brain remains an open question. One possibility are balanced excitatory-inhibitory (EI)-assemblies, which are groups of excitatory and inhibitory neurons with pronounced recurrent low-rank connectivity (Miehl et al., 2022; Mastrogiuseppe & Ostojic, 2018). Such EI-assemblies are consistent with the irregular and asynchronous spiking commonly observed in cortical circuits, potentially resulting from balanced inputs (Vogels et al., 2005; Rupprecht & Friedrich, 2018) and support the efficient representation of analog signals (Deneve & Machens, 2016). Yet it is unclear whether efficiency is the only advantage of EI-assemblies and whether they emerge naturally through optimization of an objective function. As EI-assemblies enable a continuous representation of sensory input that is focused toward learned stimuli (Meissner-Bernard et al., 2025), we wondered if they could provide a flexible and efficient representation suitable for a range of different downstream tasks. To investigate this hypothesis, we implemented spiking neural network (SNN) models and trained them using surrogate gradient techniques (Neftci et al., 2019) for tasks with varying levels of complexity and resource constraints. We then examined the changes in connectivity before and after training. To this end, we identified assemblies as community structures using the Louvain method (Blondel et al., 2008) and quantified them using the modularity metric (Newman, 2006). We found that SNNs trained on simple classification tasks without resource constraints did not exhibit stronger community structures after learning. However, networks trained on more complex tasks under resource constraints exhibited increased modularity. This effect is even more pronounced when the input connections are not plastic. Our results suggest that connectivity structures consistent with EI-assemblies could arise from network optimization under resource constraints. In future, it would be interesting to characterize signatures of EI-assemblies in trained SNNs and further investigate their putative functions, such as flexible adaptation of representations to tasks.