Spatio-temporal Structure of Excitation and Inhibition Emerges in Spiking Neural Networks with and without Biologically Plausible Constraints

We present a Spiking Neural Network (SNN) model that incorporates learnable synaptic delays using Dilated Convolution with Learnable Spacings (DCLS). We train this model on the Raw Heidelberg Digits keyword spotting benchmark using Backpropagation Through Time with surrogate gradients. Analysing the spatio-temporal structure of synaptic interactions in the network we observe that after training excitation and inhibition are grouped together both in space and time. To further enhance the efficiency and biological realism of our model, we implemented a dynamic pruning strategy that combines DEEP R for connection removal and RigL for connection reintroduction, ensuring that the network maintains optimal connectivity throughout training. Additionally, we incorporated Dale's Principle, enforcing each neuron to be exclusively excitatory or inhibitory -- aligning our model closer to biological neural networks. We observed that, after training, the spatio-temporal patterns of excitation and inhibition appeared in the more biologically plausible model as well. Our research demonstrates the potential of integrating learnable delays, dynamic pruning, and biological constraints to develop efficient SNN models for temporal data processing. Furthermore, our results enhance the understanding of spatio-temporal dynamics in SNNs -- suggesting that the spatio-temporal features which emerge from training are robust to both pruning and rewiring processes -- providing a solid foundation for future work in neuromorphic computing applications.