ºÝºÝߣs of my talk (also available for download) presented in various circumstances, including the forthcoming SOTIF safe ML workshop on Mar 28th, 2019. In this talk, I summarize my personal reflection towards dependable NN by considering a structural approach (GSN) to argue the quality with NN-specific dependability metrics. I then give an overview on our initial work in (1) estimating data completeness while admitting combinatorial explosion, (2) formal verification of neural networks, and (3) runtime monitoring for neural networks. In particular, I argue that runtime monitoring is needed even when the trained network is perfect on the training and testing sets. https://sotif-conference.iqpc.de/speakers/chih-hong-cheng