A new saliency measure for inputs selection and node pruning in neural network

This paper deals with a new saliency measure for ranking and removing the less important inputs and hidden nodes. This new metric is the result of a global sensitivity analysis, EFAST, performed on the neural network. EFAST is model independent, does not interact with the training stage and does not rely on any assumption regards to local minima for instance, contrary to a wide range of local sensitivity-based saliency measure. EFAST apportions the output variance among all the units, and hence, allows their quantitative ranking. New input selection and node pruning algorithms have been derived and are presented here. Some experimental results are provided and show with a good agreement the efficiency of the approach for inputs selection, system identification and node pruning applications.