EDR-SNN: An Energy-Efficient and Dynamic Inference Framework for Robust Spiking Neural Networks

Spiking Neural Networks (SNNs) provide an effective solution for low-power, real-time image processing on embedded and edge devices through their event-driven computing paradigm. However, their broad deployment is hindered by challenges in training efficiency, scalability, and adaptability in dynamic, resource-constrained environments. In this work, we present EDR-SNN, an energy-efficient dynamic inference framework specifically designed for embedded visual computing. At its core is the Integer Dynamic Integrate-and-Fire (IDIF) neuron model, which enhances temporal feature encoding via adaptive threshold modulation and nonlinear membrane potential updates. To facilitate efficient execution, EDR-SNN utilizes depthwise separable convolutions, reducing parameter count to 12.7\% of conventional networks and significantly lowering computational and memory requirements. Additionally, we propose OffsetReLU, a spike-compatible activation function with a tunable bias coefficient that improves noise robustness and stabilizes gradient flow. Extensive experiments on standard image classification benchmarks show that EDR-SNN achieves accuracy rates of 99.69% on MNIST, 92.96% on Fashion-MNIST, 97.66% on CIFAR-10, and 83.15% on CIFAR-100, while reducing energy consumption by 35.3% compared to state-of-the-art SNN models. These results demonstrate the effectiveness and practical potential of EDR-SNN for latency-sensitive embedded vision applications, such as autonomous drones and smart cameras.