Neural Network Repair with Reachability Analysis

Safety is a critical concern for the next generation of autonomy that likely to rely heavily on deep neural networks perception and control. This paper proposes method repair unsafe ReLU DNNs in safety-critical systems using reachability analysis. Our uses analysis calculate reachable domain DNN, then novel loss function construct its distance safe during retraining process. Since subtle changes DNN parameters can cause unexpected performance degradation, we also present minimal approach where deviation minimized. Furthermore, explore applications our agents reinforcement learning (DRL) with seamless integration algorithms. evaluated ACAS Xu benchmark rocket lander system against state-of-the-art ART. Experimental results show generate provably multiple safety specifications negligible even absence training data (Code available online at https://github.com/Shaddadi/veritex.git ).