Perturbation techniques make small changes to inputs to a model or algorithm and the. track how they change the output. They are black-box methods in that they do not need to know anything about the internal workings of the model or algorithm. Perturbatuon techniques can be used to measure the {[sensitivity}} of a model to changes and hence the uncertainty of the model. They can also be really useful for explainable AI, particularly when applied to provide a ({local explanation}} for a specific decision.
Defined on page 520
Used on Chap. 20: pages 503, 510; Chap. 21: pages 520, 521, 524, 525
Also known as perturbation