A Frobenius approximation reduction method (FARM) for determining optimal number of hidden units

A least-square approximation method is proposed to reduce the number of hidden units of a trained multilayer perceptron artificial neural network structure. In this method, the hidden neurons that contribute the most to the net function of the output layer are retained while the hidden units that contribute the least are removed. It is shown theoretically that the proposed method minimizes the Frobenius norm of the approximation error, hence the name Frobenius approximation reduction method. Also reported are simulation results on ECG classifications. The results support the theoretical predictions arid yield very encouraging performances