Enhancing the robustness of a feedforward neural network in the presence of missing data

Statistical methods applied to real-world problems account for data that is known to be missing. In contrast, neural network designers often effectively ignore missing data, assigning zero or some other constant value, and letting the well-known robustness of the network handle it. The authors propose a novel technique which greatly enhances the correct decision rate for their given example. This scheme, which does not require prohibitive computational overhead, derives substitute values for the missing ones when their existence is known.<>