Contour preserving classification for maximal reliability
Thitipong Tanprasert, Chularat Tanprasert, Chidchanok Lursinsap · 2002
This paper demonstrates that the robustness and weight fault tolerance of a neural network trained to learn a linearly separable problem can be enhanced if the network classifies the problem nonlinearly. However, a multilayer perceptron type network trained by an existing learning algorithm can normally promote the linear separability of the problem, resulting in nonoptimal solution in terms of robustness and fault tolerance. In this paper, the technique for forcing the network to optimize its internal representation towards robustness and fault tolerance is presented. The technique introduces a concept of "outpost" vectors for hiding the unwanted linearly separable characteristics of problem. Since such a task is rather specific, the "outpost" vectors are deterministically determined rather than randomized.