In general you don't actually need crazy precision to train the nets, and a small number of Taylor expansion terms tends to approximate functions fairly well anyway.