Subinterval Sensitivity for High Dimensional Models

This paper introduces an interval-based non-probabilistic sensitivity analysis method, named subinterval sensitivity. A powerful, reliable and rigorous sensitivity analysis method, which is best suited to quantify the importance of inputs purely with respect to their mathematical model. The method has only recently and partially appeared in the literature, while its scalability to high-dimensional models is claimed here for the first time. We apply subinterval sensitivity to quantify and rank the importance of the parameters of a trained neural network model while drawing comparisons with the established Sobol' sensitivity analysis method. Sensitivities on the parameters of a trained neural network can shed light on overparametrization and explainability of the neural network surrogate model.

Keywords: Interval computation, Subinterval reconstitution, Sensitivity analysis, High-dimensional models.