Quantifying How Training Gradient Sparsity Affect Spiking Neural Network Accuracy And Robustness
This research investigates how gradient sparsity affects accuracy and robustness in Spiking Neural Networks (SNNs). The study found that natural gradient sparsity provides inherent defense against adversarial attacks, allowing SNNs to achieve top performance without explicit regularization. However, this same sparsity reduces clean accuracy. The work characterizes two types of sparsity—architectural and natural—to guide future design.