Learning Efficient Representations for Sequence Retrieval

Background In many domains, the most natural representation for data is as sequences of feature vectors. For example, in speech recognition, recorded utterances are typically transformed into series of vectors which describe the frequency content over short periods of time [1]. Similarly, in natural language processing tasks, sentences are often represented as sequences of vectors where each word corresponds to a unique vector [2]. Many off-the-shelf machine learning approaches assume that feature vectors are independent, so modeling the sequential nature of these representations often necessitates special treatment.