Visualization of neural-network gaps based on error analysis

Presented in this paper is a new methodology for detection of neural-network gaps (NNG's) based on error analysis and the visualization that is applicable for n-dimensional I/O domain. The generalization problem in artificial neural networks (ANN) training is analyzed and the concept of NNG's is introduced. The NNG's are highly undesirable in ANN generalization and methods for detecting, analyzing, and eliminating them are necessary. Previous methods for NNG detection, based on two-dimensional (2-D) and three-dimensional (3-D) visualization, were not applicable for ANN's with more than three inputs. Experiments demonstrate advantages of this new methodology, which allows better understanding of the NNG phenomena using a quantitative approach.