2/2 Training time optimization for balanced accuracy/complexity neural network models

Accuracy, complexity and computational cost are very important characteristics of a model. In this paper, a dedicated neural network design and a computational cost reduction approach are proposed in order to improve the balance between the quality and computational cost of black box non linear system identification models. The proposed architecture helps to reduce the number of parameters of the model after the training phase preserving the estimation accuracy of the non reduced model. Here, we focus on the fact that this particular design helps to reduce the computational cost required for the training phase. To validate the proposed approach, we identified the Wiener-Hammerstein benchmark nonlinear system proposed in SYSID2009 [1].