On ensemble techniques of weight-constrained neural networks