Data-driven performance metrics for neural network learning

Effectiveness of data-driven neural learning in terms of both local mimima trapping and convergence rate is addressed. Such issues are investigated in a case study involving the training of one-hidden-layer feedforward neural networks with the extended Kalman filter, which reduces the search for the optimal network parameters to a state estimation problem, as compared to descent-based methods. In this respect, the performances of the training are assessed by using the Cramer-Rao bound, along with a novel metric based on an empirical criterion to evaluate robustness with respect to local minima trapping. Numerical results are provided to illustrate the performances of the training based on the extended Kalman filter in comparison with gradient-based learning.