Spiking Neural Network for Energy EfficientNeuromorphic Computing : A Hybrid ANN–SNN Model for Event Based Data on Edge Devices

Sammanfattning

Recent advances in Artificial Intelligence have led to the widespread adoption of deep ArtificialNeural Networks (ANNs) for vision-based tasks. However, these models typically require significantcomputational resources and power, making them less suitable for edge deployment. SpikingNeural Networks (SNNs), inspired by biological neural systems, offer a promising alternative dueto their event-driven and energy-efficient nature. Despite these advantages, SNNs often sufferfrom lower accuracy and difficulties in learning from sparse, asynchronous event-based data.This thesis investigates a hybrid ANN–SNN architecture designed to combine the strengths ofboth paradigms for event-based vision processing. The proposed approach integrates spatialfeature representations from ANNs with the temporal dynamics of SNNs by injecting ANNderivedmodulation signals into spiking layers. This hybrid mechanism enhances spike activityand stabilizes learning, addressing the limitations of purely SNN-based models. In addition, aknowledge distillation framework is explored, where a pre-trained ANN teacher network guidesthe training of the SNN-based student model, improving feature representation and reducingthe need for large labeled datasets.The models are evaluated using event-based datasets under varying temporal resolutions andspike threshold configurations. Experimental results demonstrate that the hybrid ANN–SNNapproach significantly improves reconstruction quality. A spike threshold of 1.0 is found toprovide the optimal balance between spike sparsity and information retention, while increasingtemporal timeframes improves performance up to a saturation point. The best configurationachieves high reconstruction accuracy with moderate spike activity, highlighting an effectivetrade-off between performance and efficiency.Furthermore, the proposed framework reduces computational overhead by limiting the need forfull backpropagation through deep ANN layers, particularly in the knowledge distillation setupwhere the teacher network remains frozen. Combined with the sparse, event-driven processingof SNNs, this makes the approach well-suited for resource-constrained edge AI applications.Finally, the thesis outlines the potential integration of the hybrid model within a System-of-Systems architecture using the Eclipse Arrowhead Framework, enabling scalable, secure, andfault-tolerant deployment in real-world distributed environments. Overall, this work demonstratesthat hybrid ANN–SNN models provide a practical pathway toward efficient and high performing neuromorphic computing systems.