On complexity of algorithms for neural network gap detection

In this paper we analyze the complexity of neural network gap (NNG) detection algorithm. The algorithm is based on visualization of error histograms after the training of an artificial neural network. For an n-dimensional neural network, Dⁿ testing points are needed for histogram creation, where D is the resolution of the domain's discretization for every input dimension. Using information about weight factors from the trained ANN, the exponential complexity of the NNG algorithm could be reduced. This reduction is due to reducing the number of testing points that are candidates for the NNG points. The results of the complexity analysis given in the paper are incorporated into new procedures that will enhance the existing algorithm for the NNG detection and reduce its complexity.