Exploiting Data Parallelism for Efficient Classification of Multi-Dimensional Patterns

Abstract

Classification of massive multidimensional data requires high computational power and memory resources. In this paper the problem of parallel implementation of the ensemble composed of classifiers operating with multi-dimensional patterns is discussed. The member classifiers of the ensemble operate in the orthogonal subspaces obtained with the Higher-Order Singular Value Decomposition (HOSVD) of the prototype pattern tensors. Computation speed-up was gained firstly by the two-level data parallelism and, secondly, thanks to application of the recently developed algorithm for fast HOSVD decomposition. In this paper we summarize our improvements to the HOSVD classifier, focusing mostly on the computational efficiency. This is achieved thanks to application of the faster decomposition algorithm and due to parallelization. First level of data parallelization is obtained thanks to separate classification of each class of the input patterns. Second level is obtained by partitioning the input dataset into smaller chunks of data, uniformly drawn from all the classes. Thanks to this, an ensemble of classifiers is obtained which allows for parallel operation both, in the training, but also in the response, modes. This allowed for computational speed-up and lower memory requirements, but also increased accuracy due to application of the ensemble instead of a single classifier. The experiments on recognition of visual objects in digital signals show high accuracy and accelerated performance during training and system response.