A spatiotemporal vector quantizer for missing sample reconstruction

In this paper we present the application of a spatiotemporal memory motivated by reaction diffusion mechanisms for missing sample reconstruction. The model is a spatiotemporal vector quantizer trained to learn and recall sequences that have a temporal order. This vector quantization scheme has an interesting property of creating time varying Voronoi cells that not only cluster feature vectors based on spatial proximity but also can provide information of the next anticipating cluster with a certain probability. This can be exploited to develop a scheme for predicting a temporal sequence. Unlike conventional prediction models which estimate the desired sample based on a linear combination of a few past samples and local statistics, this scheme employs a nonlinear memory structure to predict samples iteratively. The method is interesting in that it can predict samples missing in a burst.