ROBUST PRUNED NEURAL NETWORKS VIA ADVERSARIAL TRAINING

摘要

Systems, methods, and computer readable media are described to train acompressedneural network with high robustness. The neural network is first adversariallypre-trainedwith both original data as well as data perturbed by adversarial attacks forsome epochs,then "unimportant" weights or filters are pruned through criteria based ontheirmagnitudes or other method (e.g., Taylor approximation of the loss function),and thepruned neural network is retrained with both clean and perturbed data for moreepochs.