A discussion of migration of common neural network regularization methods on SNNs

SNN models offer notable advantages in terms of memory efficiency and energy consumption when compared to traditional artificial neural network models. However, their unique impulse propagation characteristics often result in lower prediction accuracy and limited generalization abilities. Consequently, SNNs are not widely adopted in many application domains. In light of this current state of affairs, we have undertaken an investigation into the application of various dropout techniques, including standard dropout, dropout2d, Feature dropout, and alpha dropout, which are commonly employed in traditional artificial neural networks. Our objective is to assess the impact of these dropout techniques on the training loss and test accuracy of SNN models. To conduct our experiments, we have focused on fully connected SNN models using two benchmark datasets: MNIST and CIFAR-10. Our findings indicate that SNNs subjected to multiple dropout techniques exhibit subpar performance in both training loss and test accuracy evaluations. This suggests that the straightforward application of dropout and its variants does not yield the desired improvement in the performance of SNN models.