Method of using MLP to identify decision surface for problems with highly uneven training examples

Summary form only given. Using the multilayer perceptron (MLP) to learn the boundary between two opposite classes of data is one of the most popular learning schemes. With the MLP, multiple dimensions of input variables can be correlated to form associations between input and output. However, data available for training are often biased, e.g., numerous normal training examples versus very scarce faulty training data. A strategy of training biased data has been developed. Following the bias training procedure, a tightly bounded hypersurface over a large number of normal training data can be constructed. With this approach, the faulty data can be identified with a near-zero failure rate. The false alarm rate can be reduced by further learning with additional normal training examples