CamDrop

Hongjun Wang,Guangrun Wang,Jia Li,Liang Lin

DOI: https://doi.org/10.1145/3357384.3357999

2019-01-01

Abstract:To force convolutional networks to explore more discriminative evidence throughout spatial regions, this paper presents a novel CamDrop to improve the conventional dropout in two aspects. First, by considering the intensity of class activation mapping (CAM) all around, CamDrop selectively abandons some specific spatial regions in predominating visual patterns at each iteration. In many classification tasks, CamDrop demonstrates its effectiveness and achieves considerable improvements on robust predictions for adversarial examples. Second, although dropout is a widely adopted technique that has been applied to regularize large models, the improvement in performance always attributes to better preventing DNN from overfitting. Here we give a new explanation of dropout from the perspective of optimization that it makes the upper bound of the magnitude of gradients much tighter, which leads to a more stable behavior of the gradients and effectively avoids neurons falling into the saturation region of the nonlinear activation, even when using high learning rates. Extensive experiments have been performed to prove the above two strengths of CamDrop.