Countlesslearningtasksrequiredealingwithsequentialdata.Imagecaptioning,speechsynthesis,andmusicgenerationallrequirethatamodelproduceoutputsthataresequences.Inotherdomains,suchastimeseriesprediction,videoanalysis,andmusicalinformationretrieval,amodelmustlearnfrominputsthataresequences.I