Simple heuristic methods for input parameters' estimation in neural networks
Igor V. Tetko, Vsevolod Yu. Tanchuk, A. I. Luik · 1994
We propose simple heuristic methods that can evaluate the relevance of input parameters after completing neural network training. Besides that, these methods allow correct computation of the inputs' contribution to each problem, when learning multiple tasks simultaneously. The estimations are done on a statistical base and are independent of learning procedures and cost functions. Our simulation on three different tasks shows that these approaches are effective.>