Sanity Checks for Explanation Uncertainty
This work addresses the problem of unreliable explanations for researchers and practitioners by providing incremental tests to assess explanation uncertainty methods.
The paper tackles the difficulty of evaluating explanation uncertainty in machine learning by proposing sanity checks using weight and data randomization tests, and demonstrates their validity and effectiveness on CIFAR10 and California Housing datasets, with Ensembles consistently passing tests with Guided Backpropagation, Integrated Gradients, and LIME explanations.
Explanations for machine learning models can be hard to interpret or be wrong. Combining an explanation method with an uncertainty estimation method produces explanation uncertainty. Evaluating explanation uncertainty is difficult. In this paper we propose sanity checks for uncertainty explanation methods, where a weight and data randomization tests are defined for explanations with uncertainty, allowing for quick tests to combinations of uncertainty and explanation methods. We experimentally show the validity and effectiveness of these tests on the CIFAR10 and California Housing datasets, noting that Ensembles seem to consistently pass both tests with Guided Backpropagation, Integrated Gradients, and LIME explanations.