Geometry and topology of decision regions are closely related with classification performance and robustness against adversarial attacks. In this paper, we use differential geometry to theoretically explore the geometrical and topological properties of decision regions produced by deep neural networks (DNNs). The goal is to obtain some geometrical and topological properties of decision boundaries for given DNN models, and provide some principled guidance to design and regularization of DNNs. First, we present the curvatures of decision boundaries in terms of network parameters, and give sufficient conditions on network parameters for producing flat or developable decision boundaries. Based on the Gauss-Bonnet-Chern theorem in differential geometry, we then propose a method to compute the Euler characteristics of compact decision boundaries, and verify it with experiments.
翻译:决策区域的几何学和地形学与分类性能和抵御对抗性攻击的稳健性密切相关。在本文件中,我们使用差别几何学从理论上探讨深神经网络产生的决策区域的几何学和地形学特性。目的是为给定的DNN模型取得决定界限的一些几何学和地形学特性,为DNN模型的设计和正规化提供一些原则性指导。首先,我们从网络参数的角度介绍决定边界的曲线,并为形成平坦或可开发的决定边界提供充分的网络参数条件。然后,根据Gauss-Bonnet-Chern的差别几何学理论,我们提出一种方法来计算决定边界的欧尔特性,并用实验加以核实。