Comparison Study of Sensitivity Definitions of Neural Networks

This paper compares the sensitivity definitions of neural networks´ output to input and weight perturbations. Based on the essence of the sensitivity definitions, the authors classify these sensitivity definitions into 3 categories: Noise-to-Signal Ratio, Geometrical property of derivative, Angle perturbation in Geometry space. The characteristics of these 3 categories of sensitivity definition are discussed respectively. It is sensible to classify these sensitivity definitions based on the essence of them for researchers can find other new sensitivity definitions of neural networks.