Evaluation of neural network models with generalized sensitivity analysis

A sensitivity analysis method for discovering characteristic features of th e input data using neural network classification models has been devised. T he sensitivity is the gradient of the neural network model response functio n, and because neural network models are nonlinear, the gradient depends on the point where it is evaluated, Two criteria are used for measuring the s ensitivity. The first criterion calculates the sensitivity or gradient of t he neural network output with respect to the average of the objects that co mprise each class. The second criterion measures the average sensitivity of the class objects. The sensitivity analysis was applied to temperature-con strained cascade correlation network models and evaluated with sets of synt hetic data and experimental mobility spectra. The neural network models wer e built using temperature-constrained cascade correlation networks (TCCCNs) . A weight constraint was devised for the output units of the network model s. This method implements weight decay with conjugate gradient training and yields more sensitive neural network models. Temperature-constrained hidde n units furnish more sensitive network models than networks without constra ints. By comparing the sensitivities of the class mean input and the mean s ensitivity for all the inputs of a class, the individual input variables ma y be assessed for linearity. If these two sensitivities for an input variab le differ by a constant factor, then that variable is modeled bq a simple l inear relationship. If the two sensitivities vary by a nonconstant scale fa ctor, then the variable is modeled by higher order functions in the network The sensitivity method was used to diagnose errors in the training data, a nd the test for linearity indicated a TCCCN architecture that had better pr edictability.