Increasing the Stability of CNNs using a Denoising Layer Regularized by Local Lipschitz Constant

Hamed H. Aghdam, Elnaz J. Heravi, Domenec Puig

Abstract: Recent studies have revealed that ConvNets are sensitive to small perturbations in the input. This can cause fatal consequences in smart cars because of instability of the road understanding module. In this paper, we propose an objective function regularized by the local Lipschitz constant and train a layer for restoring images. Our experiments on the GTSRB and the Caltech-Pedestrian datasets show that our modular approach not only increases the accuracy of the classification ConvNets on the clean datasets but it also increases the stability of the ConvNets against noise. Comparing our method with similar approaches shows that it produces more stable ConvNets while it is computationally similar or more efficient than these methods.

Conflicts: urv.cat