A Neurodevelopment-Inspired Deep Spiking Neural Network for Auditory Spatial Attention Detection using Single Trial EEG

Abstract

This study presents a deep probabilistic spiking neural network designed to extract discriminative spatiotemporal features from EEG signals associated with rightward and leftward auditory attention. The network self-organizes its synaptic weights and inter-layer connectivity through a biologically inspired learning rule shaped by probabilistic feedback inhibition and spike-based synchronization dynamics. To characterize the models behavior and identify optimal operating conditions, we systematically examined the effects of key parameters including inhibition strength, synchronization parameters, and preprocessing thresholds on network stability and classification performance. Simulation results show that feedback inhibition is essential for preventing uncontrolled synaptic growth, maintaining sparse connectivity, and enabling stable learning. Strong inhibition suppresses connectivity and weight development, whereas weak inhibition leads to excessive synaptic expansion. An intermediate inhibition regime achieves a balance between adaptability and stability, resulting in the most effective feature extraction. Output-layer neurons exhibited heterogeneous activation dynamics, reflecting diverse encoding of EEG input patterns. The excitatory firing probability was tightly modulated by the inhibition parameter, confirming its central role in shaping network responses. Synchronization parameters further influenced synaptic dynamics, producing nonlinear effects on spike coordination and weight evolution. Classification accuracy peaked at intermediate synchronization levels, revealing an optimal regime where inhibition and synchronization jointly support efficient classification of rightward and leftward EEG data. Additional analyses demonstrated that synaptic pruning substantially improves accuracy across all inhibition levels and that the model performs best when input spikes are generated using a preprocessing threshold of 0.3. Moreover, experimental results demonstrate that the developed neural network model achieves an average accuracy of 90% while utilizing only 10% of the available EEG data. Overall, the findings show that the proposed spiking neural network achieves its highest classification performance under moderate inhibition, intermediate synchronization, sparse connectivity, and appropriate pruning. These results highlight the promise of biologically inspired spiking architectures for decoding EEG-based attentional states, particularly in settings with limited training data and strong temporal structure.