On the robustness of Perceptron learning recurrent networks