Antoine Trouve

A Comparison of DAG and Mesh Topologies for Coarse-Grain Reconfigurable Array

In this paper, we address the hardware overhead of the dynamically reconfigurable functional unit (DRFU) in dynamically reconfigurable processors (DRP), in the context of low-power, embedded system-on-chips (E-SoC). We consider a tightly coupled DRP with a small, coarse-grain DRFU made of four columns of four ALUs. These are interconnected following one of the following interconnection scheme: direct acyclic graph or mesh. Given a large set of of custom instructions to map on the DRFU, we explore the simplification opportunities on the DRFU in order to reduce its hardware cost. We determine that it is possible to reduce its footprint by about 70 % with respect to the ALUs for both topologies and 50 % with respect to the interconnection between ALUs. We also provide the place and route algorithm to achieve these results. At the end of the paper we compare both topologies with respect to the hardware usage, the opportunities for simplifications and the complexity of the place and route algorithm. We conclude that the mesh topology is in all the cases the most desirable.

An intelligent annotation-based image retrieval system based on RDF descriptions

The notions of concept and instance are proposed to express the semantics of images.An image annotation model is proposed to annotate images at three levels.An intelligent ABIR system is implemented based on RDF descriptions.The problems of synonyms and homonyms are addressed in the our ABIR system.The proposed ABIR system provides a way to search with calculation. Display Omitted In this paper, we aim at improving text-based image search using Semantic Web technologies. We introduce our notions of concept and instance in order to better express the semantics of images, and present an intelligent annotation-based image retrieval system. We test our approach on the Flickr8k dataset. From the provided captions, we generate annotations at three levels (sentence, concept and instance). These annotations are stored as RDF triples and can be queried to find images. The experimental results show that using concepts and instances to annotate images flexibly can improve the intelligence of the image retrieval system: (1) with annotations at concept level, it enables to create semantic links between concepts and then addresses many challenges, such as the problems of synonyms and homonyms; (2) with annotations at instance level, it can count things (e.g., two people, three animals) or identify a same concept.

A survey on big data processing infrastructure: evolving role of FPGA

In todays commercial world, information is becoming a major economic resource thus leading to a statement - information is wealth. It is a technical challenge for computer systems in managing and analysing the large volumes of data coming from a variety of resources continuously over a period. Experts are in a mood of moving towards alternative hardware platforms for achieving high-speed data processing and analysis especially for streaming applications. In this paper: a) existing trends in big data processing and the necessary systems involved are studied by performing a survey on available platforms; b) recommended features and suitable hardware systems are proposed based on the operations involved in the processing. Investigation shows that, in combination with CPU and along with GPU, FPGA is a possible alternative. It can be a part of the heterogeneous platform featuring parallelism, pipelining and high performance for the operations involved in big data processing.

Semantic image retrieval for complex queries using a knowledge parser

In order to improve the retrieval accuracy of image retrieval systems, research focus has been shifted from designing sophisticated low-level feature extraction algorithms to combining image retrieval processing with rich semantics and knowledge-based methods. In this paper, we aim at improving text-based image retrieval for complex natural language queries by using a semantic parser (Knowledge Parser or K-Parser). From text written in natural language, the K-parser extracts a graphical semantic representation of the objects involved, their properties as well as their relations. We analyze both the image textual captions and the natural language queries with the K-parser. As a technical solution, we leverage RDF in two ways: first, we store the parsed image captions as RDF triples; second, we translate image queries into SPARQL queries. When applied to the Flickr8k dataset with a set of 16 custom queries, we notice that the K-parser exhibits some biases that negatively affect the accuracy of the queries. We propose two techniques to address the weaknesses: (1) we introduce a set of rules to transform the output of K-parser and fix some basic, recurrent parsing mistakes that occur on the captions of Flickr8k; (2) we leverage two popular commonsense knowledge databases, ConceptNet and WordNet, to raise the accuracy of queries on broad concepts. Using those two techniques, we can fix most of the initial retrieval errors, and accurately execute our set of 16 queries on the Flickr8k dataset.

Exploring the Importance of Negative Links through the European Parliament Social Graph

In this paper, we study the informative value of negative links in social networks, more specifically signed networks. We first process a selection of networks, generated from the 7th term (2009-2014), that were constructed by considering voting similarities between Members of the European Parliament. We propose a rework of Parallel Iterative Local Search; An algorithm to partition networks by means of solving the Correlation Clustering Problem, which takes in consideration the sign of the edges when generating selecting the communities for each node. Our rework reduces the termination time of the algorithm, and performs a faster exploration of the solution space, by means of limiting the number of iterations in local searches and also imposes a condition for a perturbation to be processed. Our approach reduced the termination required time to around 10% of the original approach, besides being able to reach better results when compared with a selection of community detection algorithms like Infomap and Walktrap, designed to process only positive links.

MAD7: a memory architecture simulator targeted at design space exploration

We introduce MAD7, a tool that rapidly simulates many-core memory architectures at a functional level. MAD7 focuses on tracking access patterns and data spatial localities rather than enforcing any precise on-chip arbitration protocols. Although not cycle accurate by nature, it provides useful insights when comparing different memory architectures under real-world workloads. Potential cache access and on-chip network usage bottlenecks can be easily spotted visually thanks to figures generated by our tool. MAD7 simulations are multi-threaded and are typically up to two orders of magnitude faster than those of a fully cycle accurate simulator.

Using Machine Learning in Order to Improve Automatic SIMD Instruction Generation

Abstract Basic block vectorization consists in extracting instruction level parallelism inside basic blocks in order to generate SIMD instructions and thus speedup data processing. It is however a double-edged technique, because the vectorized program may actually be slower than the original one. Therefore, it would be useful to predict beforehand whether or not vectorization could actually produce any speedup. In this article, we propose to do so by using a machine learning technique called support vector machine. We consider a benchmark suite containing 151 loops, unrolled with factors ranging from 1 to 20. We do our prediction offline after as well as before unrolling. Our contribution is threefold. First, we manage to predict correctly the profitability of vectorization for 70% of the programs in both cases. Second, we propose a list of static software characteristics that successfully describe our benchmark with respect to our goal. Finally, we determine that machine learning makes it possible to significantly improve the quality of the code generated by Intel Compiler, with speedups up to 2.2 times.

A Concise Conversion Model for Improving the RDF Expression of ConceptNet Knowledge Base

With the explosive growth of information on the Web, Semantic Web and related technologies such as linked data and commonsense knowledge bases, have been introduced. ConceptNet is a commonsense knowledge base, which is available for public use in CSV and JSON format; it provides a semantic graph that describes general human knowledge and how it is expressed in natural language. Recently, an RDF presentation of ConceptNet called ConceptRDF has been proposed for better use in different fields; however, it has some problems (e.g., information of concepts is sometimes misexpressed) caused by the improper conversion model. In this paper, we propose a concise conversion model to improve the RDF expression of ConceptNet. We convert the ConceptNet into RDF format and perform some experiments with the conversion results. The experimental results show that our conversion model can fully express the information of ConceptNet, which is suitable for developing many intelligent applications.

An introduction to question answering with ConceptRDF

With the development of information technologies, a great amount of semantic data is being generated on the web. Consequently, finding efficient ways of accessing this data becomes more and more important. Question answering is a good compromise between intuitiveness and expressivity, which has attracted the attention of researchers from different communities. In this paper, we propose an intelligent questing answering system for answering questions about concepts. It is based on ConceptRDF, which is an RDF presentation of the ConceptNet knowledge base. We use it as a knowledge base for answering questions. Our experimental results show that our approach is promising: it can answer questions about concepts at a satisfactory level of accuracy (reaches 94.5%).

Interactive Visualization of Quantitative Data with G2D3

This article introduces G2D3, an implementation of the grammar of graphics in JavaScript, along with two practical use cases that illustrate its practicability. It makes it possible to generate interactive visualization of quantitative data in HTML/SVG. Compared to traditional static data visualization systems such as those featured in Excel and R, G2D3 makes it possible to greatly enhance the amount of conveyed information by means of animation and interaction. Compared to other JavaScript plotting libraries such as Raphaël and D3, G2D3 leverages the expressiveness and the flexibility of the grammar of graphics to concisely generate complex visualization with many plotting dimensions, including time. It makes it possible to create a wide range of graphics with a few lines of code. G2D3 is open source and hosted in GitHub: join us!