(French version below)

You are cordially invited to attend the thesis defense of Fabien André that will be held Friday November 25, 2016 at 9:30 in room Petri-Turing and for a drink in room Minquiers.

Jury members:
Peter Triantafillou, Professor, University of Glasgow
Gaël Thomas, Professor, Telecom SudParis
Nicolas Le Scouarnec, Senior Scientist, Technicolor
Achour Mostefaoui, Professor, Université de Nantes
Anne-Marie Kermarrec, Supervisor, Mediego

Title: Exploiting Modern Hardware for Large Scale Nearest Neighbor Search

Abstract:
Many multimedia information retrieval or machine learning problems require efficient high-dimensional nearest neighbor search techniques. For instance, multimedia objects (images, music or videos) can be represented by high-dimensional feature vectors. Finding two similar multimedia objects then comes down to finding two objects that have similar feature vectors. In the current context of mass use of social networks, large scale multimedia databases or large scale machine learning applications are more and more common, calling for efficient nearest neighbor search approaches.
This thesis builds on product quantization, an efficient nearest neighbor search technique that compresses high-dimensional vectors into short codes. This makes it possible to store very large databases entirely in RAM, enabling low response times. We propose several contributions that exploit the capabilities of modern CPUs, especially SIMD and the cache hierarchy, to further decrease response times offered by product quantization.


Vous êtes cordialement invités à venir assister à la soutenance de thèse de Fabien André qui se tiendra vendredi 25 novembre à 9h30 en salle Petri-Turing ainsi qu’au pot qui suivra en salle Minquiers.

Composition du jury :
Peter Triantafillou, Professeur, University of Glasgow
Gaël Thomas, Professeur, Telecom SudParis
Nicolas Le Scouarnec, Chargé de recherche, Technicolor
Achour Mostefaoui, Professeur, Université de Nantes
Anne-Marie Kermarrec, Directrice de thèse, Mediego

Titre : Exploitation du matériel moderne pour la recherche de plus proche voisin à large échelle

Résumé :
De nombreux problèmes de recherche d’information multimédia ou d’apprentissage automatique nécessitent des techniques de recherche de plus proche voisin en haute dimensionnalité efficaces. Par exemple, les objets multimédia (images, musique ou vidéos) peuvent être représentés par des vecteurs caractéristiques de haute dimensionnalité. Trouver deux objets multimédia similaires revient alors à trouver deux objets multimédia ayant des vecteurs caractéristiques similaires. Dans le contexte actuel d’utilisation massive des réseaux sociaux, les bases de données multimédia à large échelle ou les applications d’apprentissage automatisé à large échelle sont de plus en plus courantes, exacerbant la nécessité de techniques efficaces de recherche de plus proche voisin.
Cette thèse s’appuie sur la quantification produit, une technique de recherche de plus proche voisin efficace qui compresse les vecteurs de haute dimensionnalité en codes compacts. Ceci permet de stocker des bases de données volumineuses entièrement en RAM, permettant alors des temps de réponse bas. Nous proposons plusieurs contributions exploitant les capacités des CPU modernes, notamment le SIMD et la hiérarchie de caches, pour réduire davantage les temps de réponse offerts par la quantification produit.

We are receiving Fabio Costa from the University of Goiás. He will give a talk on Wednesday, September 14 at 13:40 in room Bréhat (C001, yellow level).

Title: QoS-aware Service Selection and Cloud-based Resource Allocation for Service Choreographies

Abstract:
Service choreographies and service compositions in general rely on the proper choice of service implementations, service providers, and resource providers in order to fulfill their requirements. Besides service functionality, this choice must based on non-functional criteria, such as quality of service and efficiency of resource usage. In this talk, I will cover ongoing research on service selection and dynamic cloud resource allocation as part of an approach for the deployment and enactment of multiple service choreographies. The approach is based on the selection of the most appropriate service to fulfill each choreography role, taking into account the possibility of sharing services among different choreographies. The result of service selection is then used as the basis for QoS-aware resource allocation using multiple cloud providers, including private and public clouds. I will present current results, followed by ongoing research and future developments.

Short Bio:
Fabio M. Costa is an associate professor at the Institute of Informatics at the Federal University of Goiás, Brazil. He got his Ph.D. in Computing from Lancaster University in 2001, where worked as a member of the Distributed Multimedia Research Group. His research interests are in the area of adaptive distributed systems in general and on reflective middleware in particular, adopting concepts from model-driven engineering, notably models@runtime, in order to make reflection more tractable. Results of this research have been applied in areas such as cloud computing, smart grid, communication platforms, and ubiquitous computing, with a number of related recent journal and conference publications.

Fabio’s personal page: http://www.inf.ufg.br/~fmc/Fabio_Costa

The 6th workshop on big data and analytics (WOS6), co-organized by Technicolor and Inria, will take place at Technicolor Rennes on Thursday, the 24th of November 2016.

For more details about the event, such as the registration and schedule, please refer to the workshop page.

We are receiving Vincent Gramoli from the University of Sydney. He will give a talk on Tuesday, August 30 at 10:30 in room Sardaigne (F102, blue level).

Title: On the Danger of Private Blockchains

Abstract:
Consensus is a fundamental problem of distributed computing. While this problem has been known to be unsolvable since 1985, existing protocols were designed these past three decades to solve consensus under various assumptions. Today, with the recent advent of blockchains, new consensus implementations were proposed to make replicas reach an agreement on the order of transactions updating what is often referred to as a distributed ledger. Very little work has however been devoted to explore its theoretical ramifications. As a result, it is often unclear whether the same systems could be adapted to work in different environments.
In this position paper, we explore the use of the Ethereum blockchain protocol in the context of a private chain where the set of participants is controlled. We argue that foundations are needed in order to precisely capture the guarantees of the consensus protocols of novel blockchain systems before one can deploy them safely. To this end, we define the termination of consensus to characterize when blockchain transactions commit and describe the existence of the Blockchain Anomaly in existing proof-of-work private chains.

Bio:
Vincent Gramoli is an academic at the University of Sydney and a senior researcher at Data61-CSIRO. His research interest is in distributed computing and his work on blockchain focuses on the consensus problem. Vincent was affiliated with INRIA, Cornell, Neuchatel and EPFL in the past, and received his PhD from University of Rennes and his Habilitation from UPMC Sorbonne University.

Vincent’s personal page: http://poseidon.it.usyd.edu.au/~gramoli

Our paper “Speed for the elite, consistency for the masses: differentiating eventual consistency in large-scale distributed systems” by Davide Frey, Achour Mostefaoui, Matthieu Perrin, Pierre-Louis Roman and François Taïani has been accepted at the 35th IEEE Symposium on Reliable Distributed Systems (SRDS 2016).

« Newer Posts - Older Posts »