In this paper, the possibility of using an orthogonal basis to train a collection of artificial neural networks in a face recognition task is discussed. This orthonormal basis is selected from a dictionary of orthonormal bases consisting of wavelet packets. Here, a basis is obtained by maximizing a certain discriminant measure among classes of training images. Once such a basis is selected, its basis vectors are ordered according to their power of discrimination and the first N most local discriminant basis vectors are retained for image decomposition purpose. By projecting all training images onto an individual basis vector of these N most discriminant basis vectors, N versions of the training set at different spatial/scale resolutions are then created. Next, N multilayer feed- forward neural networks are trained independently by N different resolution-specific training sets. After networks have been trained, they are combined to form an ensemble of networks. Our proposed method takes advantage of the fact that the dimensionality of the pattern recognition problem at hand is reduced, but the important information is still contained, and at the same time, some correlations between neighboring inputs are included. Furthermore, the performance of our proposed network is improved over a single neural network as a result of the ensemble and the nonlinear property of neural networks. Finally, this method is applied to a face recognition task using the Yale Face Database. From the experimental results , the performance of our method is better than a conventional back-propagation network and a wavelet packet parallel consensual neural network in terms of both computation and generalization ability.