Sensitivity analysis uses small {peturbations}} of input data to see how much the output of an algorithm changes. This can be used as part of the system design process, if the output is highly sensitive to certain input fields then it is impirtant that these are collected or measured with high accuracy. Sensitvity analysis is also a central part of several techniques in explainable AI. For example, consider a black box image classifiaction system that uses a deep neurak network; by measuring the sensitivity of the classification of an image to each pixle one can createa hear map showing which portions of the image are being used by the network to classify the image.
Defined on page 520
Used on pages 520, 524