Robustness quantification of recurrent neural network using unscented transform

Abstract

Artificial recurrent neural network has been proved to be a valuable tool in modeling nonlinear dynamical systems. Robustness study of recurrent neural network is critical to its successful implementation. The goal of robustness study is to reduce the sensitivity of modeling capability to parametric uncertainties or make the network fault tolerant. In this study, an uncertainty propagation analysis is performed to quantify the robustness of a recurrent neural network output due to perturbations in its trained weights. An uncertainty propagation analysis-based robustness measure has been proposed accordingly and further compared with available performance loss-based and sensitivity matrix-based approaches. Results show that the proposed robustness measure approach is more efficient, generic, and flexible to quantify the robustness of a recurrent neural network.