AbstractTimeseriesoftenhaveatemporalhierarchy,withinformationthatisspreadoutovermultipletimescales.Commonrecurrentneuralnetworks,however,donotexplicitlyaccommodatesuchahierarchy,andmostresearchonthemhasbeenfocusingontrainingalgorithmsratherthanontheirbasicarchitecture.Inthispa-perwe