Empirical studies of separability and prediction using threshold logic units

摘要

Several properties of threshold logic units (TLU's) were studied by actual training on vectors filled with random numbers (a Gaussian distribution). Trends have been observed relating predictive ability to the training set size (N)/dimensionality (D) ratio and also linking the degree of category set overlap to the frequency with which subsets are successfully trained. The results further suggest that predictive ability may not change significantly over wide ranges of D providing the N/D ratio is constant and greater than two or three.