The Impact of Internal Noise on Deep and Spiking Neural Networks

This work investigates the impact of internal noise on both deep fully connected artificial neural networks (ANNs) and spiking neural networks (SNNs) implemented with FitzHugh–Nagumo systems. For deep networks, simulations show that noise in earlier hidden layers degrades classification accuracy more severely, while significant performance loss occurs only for noise intensities exceeding approximately 10% of the neuron output range. In SNNs, the effect of noise strongly depends on its type and location: additive noise in the first equation has minimal impact, whereas multiplicative noise can cause a sharp performance drop once it exceeds experimentally relevant levels. These results underscore the importance of network architecture and neuron dynamics in determining the noise robust neural networks.