Abstract: An enhanced layer wise learning algorithm for training multilayer feedforward neural networks is proposed. The output function of hidden neural units is determined by the samples of the given system adaptively. The weights of the input layer to the hidden layer are first supposed to be invariant, then the weights of the current layer are then modulated. The output values of the forward layer, which are used as temporal signals for training the weights of the forward layer, are then estimated. The calculation of the weights of every layer and the estimation of the error of the hidden layer are changed into least squares problems, and the process goes on layer by layer, up to the input layer. The results of digital simulations and applications show the effectiveness of the algorithm.