@Techreport{MR11tech,
  AUTHOR =   {Iveta Mr{\'a}zov{\'a} and Zuzana Reitermanov{\'a}},
  TITLE =    {A New Sensitivity-Based Feature Selection Technique for Feed-Forward Neural Networks That Improves Generalization},
  INSTITUTION =  {Charles University, Faculty of Mathematics and Physics},
  YEAR =     {2011},
  ADDRESS =  {Prague},
  URL =      {http://popelka.ms.mff.cuni.cz/cerno/files/mrazova_reitermanova_feature_selection.pdf},
  ABSTRACT = {Multi-layer neural networks of the back-propagation type became already a well-established tool
        used successfully in various application areas. Efficient solutions to complex tasks currently dealt
        with obviously require sufficient generalization capabilities of the formed networks and an easy interpretation
        of their function. For this reason, we will introduce here a new feature selection technique
        called SCGSIR inspired by the fast method of scaled conjugate gradients (SCG) and sensitivity analysis.
        Enforced internal knowledge representation supports an easy interpretation of the formed network
        structure. Network sensitivity inhibited during training impacts successful pruning of input
        neurons and optimization of network structure, too. Experiments performed so far on the problem
        of binary addition and on real data obtained from the World Bank yield promising results outperforming
        reference techniques when considering both their ability to find networks with and optimum
        architectures and generalization capabilities of the trained networks.},
  KEYWORDS = {neural networks, feature selection},
}