Training of optimal cluster separation networks
Andreas Wendemuth · Journal of Physics A Mathematical and General · 1994
Finding the optimal separation of two clusters of normalized vectors corresponds to training thresholds and weights in a neural network of maximum stability. In order to achieve this, two local iterative algorithms are presented which treat threshold and weights all in one, avoiding the need to calculate any intermediate 'test' quantities. Convergence is proved, and the separation/stability obtained is shown to match theoretical predictions and to be superior to existing algorithms.