Training and Analyzing Deep Recurrent Neural Networks

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