Input Ranking Revisited: A Theoretical Input Sensitivity Approach
Sanggil Kang, Can Işık, Steven Brian Morphet · 2003
In feedforward neural networks, all inputs contribute to a greater or lesser extent when calculating the outputs. Therefore, inputs may be ordered from the greatest contributor to the least. Input ranking is non-trivial — cursory examination of the weight and bias matrices fails to reveal ranking. Solving the ranking issue allows for elimination of inputs with little influence on output. This paper presents a new method of determining the input sensitivity of three-layer feedforward neural networks. Specifically, sensitivity of an input is independent of the magnitudes of the remaining inputs, providing an unambiguous ranking of input importance. Small changes to influential inputs will result in great changes to output. This concept motivated the theoretical approach to input ranking. Examination of theoretical results will demonstrate the correctness of this approach. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.