ABSTRACT
In addition to superior performance and some other appealing features, those DNN-assisted methods also have some potential limitations. In this chapter, we discuss some concerns, such as observed data being different from training data, improperly chosen input variables for DNNs, etc. We further focus on several mitigation approaches. Such a discussion reflects a key point in Chapter 1 that human intelligence is important to ensure that those AI-based methods perform properly or are able to achieve certain bottom-line properties even under some unexpected scenarios. We strongly recommend a well-designed framework to safeguard AI-based methods based on statistical or scientific expertise.
