Sensitivity analysis of neural models
D. Lamy, Pierre Borne · 2002
This paper investigates the sensitivity of neural models to weights perturbation in a system identification task. Analytical expression for sensitivity is derived from a notation based on Kronecker product and vector valued function of matrix. Experimental results highlight this sensitivity measure when investigating model structure. A comparison with statistical sensitivity results confirms usefulness of our approach. Search for minimum output sensitivity appears to be a nice indicator for proper model order choice.>