Adversarial Margin Maximization Networks

The tremendous recent success of deep neural networks (DNNs) has sparked a surge interest in understanding their predictive ability. Unlike the human visual system which is able to generalize robustly and learn with little supervision, DNNs normally require massive amount data new concepts. In addition, research works also show that are vulnerable adversarial examples-maliciously generated images seem perceptually similar natural ones but actually formed fool learning models, means models have problem generalizing unseen certain type distortions. this paper, we analyze generalization ability comprehensively attempt improve it from geometric point view. We propose margin maximization (AMM), learning-based regularization exploits an perturbation as proxy. It encourages large input space, just like support vector machines. With differentiable formulation perturbation, train regularized simply through back-propagation end-to-end manner. Experimental results on various datasets (including MNIST, CIFAR-10/100, SVHN ImageNet) different DNN architectures demonstrate superiority our method over previous state-of-the-arts. Code for reproducing will be made publicly available.