Learning Sequence Neighbourhood Metrics

Abstract

Storing short descriptors of sequential data has several benefits. First, they typically require much less memory and thus make processing of large data sets much more efficient. Second, if the descriptors are formed as vectors, e.g. x 2 Rn, numerous algorithms tailored towards static data can be applied. Instead of applying static data algorithms to dynamic data, we propose to learn a mapping from sequential data to static data first. This can be done by combining recurrent neural networks (RNNs), a pooling operation and any differentiable objective function for static data. In this work, we present how neigbourhood components analysis (NCA) (Goldberger et al. 2004) can be used to learn meaningful representations which lead to excellent classification results and visualizations on a speech dataset.