摘要
Spiking Neural Networks (SNNs) offer energyefficient computation but still face challenges such as training instability, gradient vanishing, and inefficient hardware deployment. We propose a spiking convolutional neural network (SCNN) framework that integrates three key innovations. First, a Parameterized Sigmoid Gradient with Softplus (PSGS) enables adaptive gradient shaping to enhance early stability and late convergence. Second, a Hybrid Surrogate Gradient Training (HSGT) method dynamically switches between smooth and sharp surrogate functions over time. Third, we propose a PiecewiseLinear Complementary Leaky Integrate-and-Fire (PCLIF) neuron that employs piecewise linear activation and simplified memory updates to reduce inference cost. Our model achieves 97.57% accuracy on DVS128 Gesture and 80.40% on CIFAR10-DVS using only 10 time steps. Compared to CLIF, PCLIF reduces area and power by 44.8% and 70.9%, respectively, while maintaining competitive accuracy. These results demonstrate that our co-design of training strategy and neuron architecture enables training effectiveness and efficient deployment in low-power neuromorphic systems.