Neural network verification aims to provide provable safety and robustness guarantees for models used in safety-critical applications such as aircraft control and self-driving cars. While earlier approaches before 2017 only scaled to networks with hundreds of neurons, significant progress in neural network verification has been made. Thus the 2021 International Verification of Neural Networks Competition (VNN-COMP 2021) was formed to provide an objective comparison of state-of-the-art tools in terms of scalability, flexibility, and speed.
Computer Science faculty Cho -Jui Hsieh and his students Zhouxing Shi, Yihan Wang, in collaboration with Huan Zhang (UCLA CS alumni) and teams from CMU, Columbia, Northeastern, have won VNN-COMP 2021 with the top score. Their toolbox called α,β-CROWN achieved the highest score of 779.2 out of a maximum of 800, outperforming 11 competing tools, and winning by a large margin (the 2nd and 3rd places scored 701.2 and 582.0 points, respectively). Many universities and companies participated in this competition, including researchers from Oxford, MIT, UIUC, Stanford, ETH, and Google’s DeepMind. The UCLA tool α,β-CROWN also beat the previous year’s winner ERAN, which helped launch a startup company that raised over $2.5M for building reliable AI systems.