Objective functions for training new hidden units in constructive neural networks

摘要: In this paper, we study a number of objective functions for training new hidden units in constructive algorithms multilayer feedforward networks. The aim is to derive class the computation which and corresponding weight updates can be done O(N) time, where N patterns. Moreover, even though input freezing applied during process computational efficiency, convergence property using these still preserved. We also propose few tricks that used improve optimization under practical situations. Their relative performance set two-dimensional regression problems discussed.