Computing Time Lower Bounds for Recurrent Sigmoidal Neural Networks