Efficient Training and Neuro-Encoding for Bridging Hybrid ANN and SNN Computation

The complementary strengths of Spiking Neural Networks (SNNs) and Artificial Neural Networks (ANNs) have promoted interest in leveraging hybrid ANN/SNN computation. While most existing efforts focus on ANN-SNN conversion for pure SNN inference, hybrid ANN/SNN inference present unique challenges where complexity and performance in both domains are critical. Key limitations include achieving ultra-low latency, maintaining unified training parameters for resource sharing, and developing efficient neural and encoding models for hybrid data interactions. To address these challenges, We introduce the Adaptive Clip-Floor-Shift (ACFS) activation to bridge the ANN-SNN gap with unified parameters, balancing inference accuracy and complexity across both domains. Our Hybrid Neuro-Encoding Bridge (HNEB) integrating Clipped-ReLU for ANNs, proposed Selective Integrate-and-Fire (SIF) model for enhanced SNN sparsity, and a Stateless Spike Encoding (SSE) mechanism for resource-efficient activation-spike conversion. Experimental results on VGG16 and ResNet demonstrate SNNs achieving competitive accuracy (<inline-formula><tex-math notation="LaTeX">$\leq \! 0.89\%$</tex-math></inline-formula> loss) versus ANNs at ultra-low latency (e.g., <inline-formula><tex-math notation="LaTeX">$T \leq 4$</tex-math></inline-formula> for CIFAR10, <inline-formula><tex-math notation="LaTeX">$T \leq 8$</tex-math></inline-formula> for CIFAR100). Experimental analysis reveals Hybrid Neural Netwroks (HNNs) provide superior energy-accuracy trade-offs, improving energy efficiency by up to 84.13% over pure SNNs while maintaining accuracy through layer-wise ANN/SNN partitioning and minimized encoding overhead.