SyReNN: A tool for analyzing deep neural networks

Tools and algorithms for the construction and analysis of systems: a special issue on tool papers for TACAS 2021

“…The SyReNN tool focuses on the expressive subclass of Deep Neural Networks (DNNs) in which activation functions are Piecewise Linear [7]. Matthew Sotoudeh, Zhe Tao, and Aditya V. Thakur show that a precise symbolic representation of the input/output relation of a DNN can be obtained and exploited for visualization, analysis and even repair of the DNN.…”

When to Trust AI: Advances and Challenges for Certification of Neural Networks

“…Note that the combination of the four polytopes actually covers the hidden vector space [−2, 2] × [−2, 2], and the resulting preimage polytopes on the input layer cover the region 1], which certifies that the correct decision is taken for the entire region under investigation. b) SyReNN: SyReNN is proposed in [47] to compute the symbolic representation of a neural network so as to understand and analyze its behaviours. It targets low-dimensional input subspaces and computes their exact symbolic partitioning, on which the mapping function is completely linear.…”