Studying the invertibility of deep neural networks (DNNs) provides a principled approach to better understand the behavior of these powerful models. Despite being a promising diagnostic tool, a consistent theory on their invertibility is still lacking. We derive a theoretically motivated approach to explore the preimages of ReLU-layers and mechanisms affecting the stability of the inverse. Using the developed theory, we numerically show how this approach uncovers characteristic properties of the network.
Complexity of Injectivity and Verification of ReLU Neural Networks
Technische Universität Berlin, Faculty IV Institute of Software Engineering and Theoretical Computer Science, Algorithmics and Computational Complexity（柏林技术大学，第四学院软件工程与理论计算机科学研究所，算法与计算复杂性）；Technische Universität Berlin, Faculty II, Institute of Mathematics, Combinatorial Optimization and Graph Algorithms（柏林技术大学，第二学院数学研究所，组合优化与图算法）
Gradient Networks for Universal Magnetic Modeling of Synchronous Machines
Department of Electrical Engineering and Automation, Aalto University（电气工程与自动化系，阿alto大学）；Aalto University House of AI（阿alto大学人工智能学院）；Research Council of Finland Centre of Excellence in High-Speed Electromechanical Energy Conversion Systems（芬兰研究理事会高速电磁能量转换系统卓越中心）；Aalto School of Electrical Engineering（阿alto电气工程学院）；ABB Oy, Drives（ABB Oy，驱动部门）
Approximating Positive Homogeneous Functions with Scale Invariant Neural Networks
Technical University of Munich（慕尼黑工业大学）；Holsten Systems GmbH（霍尔斯特恩系统有限公司）；Munich Center for Machine Learning（慕尼黑机器学习中心）；Munich Data Science Institute（慕尼黑数据科学研究所）
