Learning Internal Representations by Error Propagation