基于规格的神经网络简化用于可扩展形式验证（英文版）

英文标题：Specification-Driven Neural Network Reduction for Scalable Formal Verification中文摘要：本文提出了一种保守的神经网络缩减方法，该方法能够在验证减小的网络的同时，确保验证原始网络，可以适用于任何类型的激活函数，使得网络能够被缩小到不到原来的 5% 的大小，从而显著减少了验证时间。英文摘要：Formal verification of neural networks is essential before their deploymentin safety-critical settings. However, existing methods for formally verifyingneural networks are not yet scalable enough to handle practical problems thatinvolve a large number of neurons. In this work, we propose a novel approach toadd