A distributed SVM for scalable image annotation

Abstract

Machine learning techniques have facilitated image retrieval by automatically classifying and annotating images with keywords. Among them Support Vector Machines (SVMs) are used extensively due to their generalization properties. However, SVM training is notably a computationally intensive process especially when the training dataset is large. This paper presents MRSVM, a distributed SVM algorithm for large scale image annotation which partitions the training data set into smaller subsets and train SVM in parallel using a cluster of computing nodes. MRSVM is evaluated in an experimental environment showing that the distributed SVM algorithm reduces the training time significantly while maintaining a high level of accuracy in classifications.

A MapReduce‐based parallel K‐means clustering for large‐scale CIM data verification

“…Considering the total processing power of the cluster is P ¼ X n i¼1 p i where n represents the number of the processors employed in the cluster and p i represents the processing speed of the ith processor, for a Hadoop cluster with a total computing capacity P, the level of heterogeneity of the Hadoop cluster can be defined using Eq. (5).…”

“…However, with the rapid development of a variety of computer systems in the electric power grid, it has become a challenging issue to ensure a large amount of CIM data to be correct and consistent all the time.MapReduce has become a major computing model in support of data-intensive applications [4]. MapReduce facilitates a number of important functions such as partitioning the input data, scheduling MapReduce jobs across a cluster of participating nodes, handling node failures, and managing the required network communications [5]. We have implemented a MapReduce-based parallel K-means clustering for scalable information retrieval [6].…”

Multimedia Applications and Security in MapReduce: Opportunities and Challenges

“…Catanzaro et al proposed a parallel SMO algorithm based on MapReduce, but this algorithm is somewhat inefficient because of the iterative nature of SMO. In , Alham discussed an efficient and scalable approach based on a single MapReduce phase. They announced that this method has minimal data movement between nodes and that it also minimizes communication overheads.…”

Support Vector Machines for Classification

“…SMO speeds up the training phase only, with no control over the number of support vectors or testing time. To achieve additional acceleration, many parallel implementations of SMO (Zeng et al 2008;Peng, Ma, and Hong 2009;Catanzaro et al 2008;Alham et al 2010;Cao et al 2006) were developed on various parallel programming platforms, including graphics processing unit (GPU) (Catanzaro et al 2008), Hadoop MapReduce (Alham et al 2010), and message passing interface (MPI) (Cao et al 2006).…”