Fig. 4: Results of ASRNNs and Rhythm-ASRNNs against various noise perturbations.
From: Efficient and robust temporal processing with neural oscillations modulated spiking neural networks
a–d Comparison of the accuracy drop ratio of ASRNNs and Rhythm-ASRNNs under varying levels of input-related Gaussian noise and network-related noises, including thermal noise, silence noise, and quantization noise. e–h Comparison of perturbation distances for ASRNNs and Rhythm-ASRNNs across various types of noise perturbations, illustrated in (a–d). Note that the highest noise level was utilized in this analysis. The perturbation distance is quantified using the Euclidean distance between the network representations prior to and following the introduction of noise. i–l Comparison of the changes in average firing rate and average perturbation distance for ASRNNs and Rhythm-ASRNNs under various types of perturbations. Rhythm-ASRNNs with a smaller duty cycle exhibit greater robustness against noise perturbations. In the legend, `dc' represents the duty cycle of the oscillatory modulation signal used in Rhythm-ASRNNs. The numbers following the colon specify the lower and upper bounds of the initial distribution of the duty cycle. m–p Comparison of the accuracy drop ratio and perturbation distances for ASRNNs and Rhythm-ASRNNs across various types and levels of adversarial attacks. The error bars represent the standard deviation of three runs with different random seeds.
