Minimizing Worst-Case Violations of Neural Networks