Verification for Learning and Learning for Verification

Machine learning methods have proven very effective for control, planning and estimation onboard complex robotic systems. Unfortunately, the safety and reliability of the resulting system consequently becomes much more difficult to analyze. One solution to this problem is the small but growing body of work focused on formal verification of neural networks. However, many of the first works published in this area considered simple settings where the networks were analyzed in isolation, when in reality networks are used as components within large, complex systems. The work done during my PhD and postdoctoral fellowship and the work that will be done in my forthcoming new research group focuses on developing methods to formally verify neural networks used in complex settings, for example such as in feedback control systems. In this talk, I will give background on neural network verification, review recent results, and present my planned future work.