Impact of Noisy Input on Evolved Spiking Neural Networks for Neuromorphic Systems

Abstract

In this work we leverage a simple spiking neuromorphic processor and an evolutionary-based training method to train and test networks in classification and control applications with noise injection in order to explore the resilience and robustness of spiking neural networks on neuromorphic systems. Through our implementation, we were able to observe that injecting noise within the training phase produces more robust networks that are more resilient to noise within the testing phase. Compared to the performance of other popular classifiers on simple data classification tasks, SNNs perform behind nearest neighbors and linear SVM, and above decision trees and traditional neural networks, with respect to performance in the presence of input noise.