Vol. 30 · No. 5

The concept vector model generalizes standard representations of similarity concept in terms of tree-like structure. In the model, each concept node in the hierarchical tree has ancestor and descendent concept nodes composing its relevancy nodes, thus a concept node is represented as a concept vector according to its relevancy nodes' density and the similarity of the two concepts is obtained by computing cosine similarity between their vectors. In addition, the model is adjusted in terms of local density and multiple descendents problem. The model contains structure information inherent and hidden in the tree. We show that this measure compares favorably to other measures, and it is flexible in that it can make comparisons between any two concepts in a hierarchical tree without relying on additional dictionary or corpus information.

Every real object having certain properties can be described by a number of descriptors, visual or other, e.g. mechanical, chemical etc. A set of descriptors (features) characterizing a given object is described in the paper by a vector of descriptors, where each entry of the vector determines a value of some feature of the object. In general, it is important to describe the object as completely as possible, which means by a large number of descriptors. This paper deals with a problem of selection of a proper subset of descriptors, which have the most substantial influence on the properties of the object, so that irrelevant descriptors could be excluded. For this purpose, we introduce a new method, Maximum Coverage Method (MCM). This method has been combined with optimization by a classical genetic algorithm. The described method is used for a data pre-processing, with the resulting selected features serving as an input for a neural network.

To reduce the semantic gap in the content based image retrieval (CBIR) systems we propose a fuzzy rule base approach. By submitting a query to the proposed system, it first extracts its low-level features and then checks its rule base for determining the proper weight vector for its distance measure. It then uses this weight vector to determine what images are more similar to the query image. For the training purpose, an algorithm is provided by which the system adjusts its fuzzy rules' parameters by gathering the trainers' opinions on which and how much the image pairs are relevant. For further improving the performance of the system, a feature space dimensionality reduction method is also proposed. We compared the proposed method with some other common ones. Our experiments on a subset of the Corel database containing 59 600 images show that the proposed method is more precise than these compared methods based on the precision and recall criterions.

The increasing complexity of modern architectures and memory models challenges the design of optimizing compilers. It is mandatory to perform several optimizing transformations of the original program to exploit the machine to its best, especially for scientific, computational-intensive codes. Aiming at investigating the best transformation sequence and the best transformation parameters simultaneously, this paper presents a novel loop transformation framework, which integrates the advantages of polyhedral model and model-guided iterative compilation to create a powerful framework that is capable of fully automated non-parametric transformations and model-guided parametric transformations as well as automatic parameter search. The framework employs polyhedral model to facilitate the search of non-parametric code transformation composition, and designs a transformation model based on hardware performance counters to guide when, where and in what order to apply transformations to get the most benefit, finally uses Nelder-Mead simplex algorithm to find the optimal parameters. The framework is demonstrated on three typical computational kernels for code transformations to achieve performance that greatly exceeds the native compiler, and is significantly better than state-of-the-art polyhedral model based loop transformations and iterative compilation, generating efficient code on complex loop nests.

One limitation of vector-based LDA and its matrix-based extension is that they cannot deal with heteroscedastic data. In this paper, we present a novel two-dimensional feature extraction technique for face recognition which is capable of handling the heteroscedastic data in the dataset. The technique is a general form of two-dimensional linear discriminant analysis. It generalizes the interclass scatter matrix of two-dimensional LDA by applying the Chernoff distance as a measure of separation of every pair of clusters with the same index in different classes. By employing the new distance, our method can capture the discriminatory information presented in the difference of covariance matrices of different clusters in the datasets while preserving the computational simplicity of eigenvalue-based techniques. So our approach is a proper technique for high-dimensional applications such as face recognition. Experimental results on CMU-PIE, AR and AT & T face databases demonstrate the effectiveness of our method in term of classification accuracy.

The presence of a large number of component models to date should be able to offer software developers a wide variety of component models – which they can easily choose from – for their software development projects. However, the opposite situation is currently observed, where the presence of many component models has caused difficulties in making the selection. Lack of properties or characteristics that can be used as a basis to perform objective comparison between the existing models is believed to have caused the difficulties. In this paper, a list of component model properties is derived by thoroughly examining the available component model definitions. Results from a comparative analysis performed on six component models using the properties show that the properties enable a more objective comparison between the existing component models to be performed.

Most clustering algorithms assume that all the relevant data are available on a single node of a computer network. In the emerging distributed and networked knowledge environments, databases relevant for computations may reside on a number of nodes connected by a communication network. These data resources cannot be moved to other network sites due to privacy, security, and size considerations. The desired global computation must be decomposed into local computations to match the distribution of data across the network. The capability to decompose computations must be general enough to handle different distributions of data and different participating nodes in each instance of the global computation. In this paper, we present a methodology and algorithm for clustering distributed data in d-dimensional space, using nearest neighbor clustering, wherein each distributed data source is represented by an agent. Each such agent has the capability to decompose global computations into local parts, for itself and for agents at other sites. The global computation is then performed by the agent either exchanging some minimal summaries with other agents or traveling to all the sites and performing local tasks that can be done at each local site. The objective is to perform global tasks with a minimum of communication or travel by participating agents across the network.

As the momentum behind Chip Multi-Processors (CMPs) continues to grow, Last Level Cache (LLC) management becomes a crucial issue to CMPs because off-chip accesses often involve a big latency. Private cache design is distinguished by smaller local access latency, good performance isolation and easy scalability, thus is becoming an attractive design alternative for LLC of CMPs. This paper proposes Balanced Private Non-Uniform Cache Architecture (BP-NUCA), a new LLC architecture that starts from private cache design for smaller local access latency and good performance isolation, then introduces a low cost mechanism to dynamically migrate private blocks among peer private caches of LLC to improve the overall space utilization. BP-NUCA achieves this by measuring the cache access pressure level that each cache set experiences at runtime and then using the information to guide block migration among different private caches of LLC. A heavily accessed set, namely a set with high access pressure level, is allowed to migrate its evicted blocks to peer private caches, replacing blocks of sets which are with the same index and have low access pressure level. By migrating blocks from heavily accessed cache sets to less accessed cache sets, BP-NUCA effectively balances space utilization of LLC among different cores. Experimental results using a full system CMP simulator show that BP-NUCA improves the overall throughput by as much as 20.3