A neural architecture based on linear predictability is used to separate linear mixtures of signals. The architecture is divided in two parameterers groups, one modeling the linear mixture of signals and the other computing the linear predictions of the reconstructed signals. The network weights correspond to the mixing matrices and coefficients of the linear predictions, while the values computed by the network units correspond to the predicted and reconstructed signal values. A quadratic error is iteratively minimized to approximate the mixing matrix and to maximize the linear predictability. Experiments with toy and acoustic signals show the feasibility of the architecture.