Back from Supercomputing by Bruno Raffin (Datamove)
– November 28, 2019
Bruno will present his report from the new trends he saw this year at supercomputing as well as explain some technical talks he liked presented at the conference.
Phd Defense: Autonomic Resilience of Distributed IoT Applications in the Fog, by Umar Ozeer (Polaris)
– December 11, 2019
Abstract:
Recent computing trends have been advocating for more distributed paradigms, namely Fog computing, which extends the capacities of the Cloud at the edge of the network, that is close to end devices and end users in the physical world. The Fog is a key enabler of the Internet of Things (IoT) applications as it resolves some of the needs that the Cloud fails to provide such as low network latencies, privacy, QoS, and geographical requirements. For this reason, the Fog has become increasingly popular and finds application in many fields such as smart homes and cities, agriculture, healthcare, transportation, etc.
The Fog, however, is unstable because it is constituted of billions of heterogeneous devices in a dynamic ecosystem. IoT devices may regularly fail because of bulk production and cheap design. Moreover, the Fog-IoT ecosystem is cyber-physical and thus devices are subjected to external physical world conditions which increase the occurrence of failures. When failures occur in such an ecosystem, the resulting inconsistencies in the application affect the physical world by inducing hazardous and costly situations.
In this Thesis, we propose an end-to-end autonomic failure management approach for IoT applications deployed in the Fog. The proposed approach recovers from failures in a cyber-physical consistent way. Cyber-physical consistency aims at maintaining a consistent behavior of the application with respect to the physical world, as well as avoiding dangerous and costly circumstances.
The approach was validated using model checking techniques to verify important correctness properties. It was then implemented as a framework called F3ARIoT. This framework was evaluated on a smart home application. The results showed the feasibility of deploying F3ARIoT on real Fog-IoT applications as well as its good performances in regards to end user experience.
Tropical approach to semidefinite programming and mean payoff games, by Mateusz Skomra (ENS Lyon)
– December 19, 2019
Semidefinite programming (SDP) is a fundamental tool in convex and polynomial optimization. It consists in minimizing linear functions over spectrahedra (sets defined by linear matrix inequalities). In particular, SDP is a generalization of linear programming. In this talk, we discuss the nonarchimedean analogue of SDP, replacing the field of real numbers by the field of Puiseux series. Our methods rely on tropical geometry and, in particular, on the study of tropicalization of spectrahedra. We show that, under genericity conditions, tropical spectrahedra encode Shapley operators associated with stochastic mean payoff games. As a result, a large class of semidefinite feasibility problems defined over Puiseux series can be solved efficiently using combinatorial algorithms designed for stochastic games. Conversely, we use tropical spectrahedra to introduce a condition number for stochastic mean payoff games. We show that this conditioning controls the number of value iterations needed to decide whether a mean payoff game is winning. In particular, we obtain a pseudopolynomial bound for the complexity of value iteration provided that the number of random positions is fixed.
The talk is based on joint works with X. Allamigeon, S. Gaubert, and R. Katz.
Capacity of a LoRaWAN cell, by Martin Heusse (Drakkar)
– January 16, 2020
We propose a model to estimate the packet delivery rate in a LoRaWAN cell, when all nodes have the same traffic generation process and may use repetitions. The model predicts the transmission success rate for any cell range and node density, with similar traffic from all nodes. We find that the transmission success depends on striking a balance between the adverse effects of attenuation and collisions; in small cells, it is highly dependent on the suitable allocation of spreading factors, whereas using packet repetitions is more effective in large cells.
Can random matrices change the future of machine learning? by Romain Couillet (Gipsa)
– January 23, 2020
Romain COUILLET (professor at CentraleSupélec, University ParisSaclay; IDEX GSTATS Chair & MIAI LargeDATA Chair, University Grenoble-Alpes)
Title: Can random matrices change the future of machine learning?
Abstract: Many standard machine learning algorithms and intuitions are known to misbehave, if not dramatically collapse, when operated on large dimensional data. In this talk, we will show that large dimensional statistics, and particularly random matrix theory, not only can elucidate this behavior but provides a new set of tools to understand and (sometimes drastically) improve machine learning algorithms. Besides, we will show that our various theoretical findings are provably applicable to very realistic and not-so-large dimensional data.
The design of efficient algorithms is closely linked to the evaluation of their performances. My work focuses on the use of stochastic models for the performance evaluation of large distributed systems. I am interested in developing tools that can characterize the emergent behavior of such systems and improve their performance. This leads me to solve stochastic control and optimization problems, notably through operations research methods. These problems suffer from combinatorial explosion: the complexity of a problem grows exponentially with the number of objects that compose it. It is therefore necessary to design models but also algorithmic processes whose complexity does not increase too rapidly with the size of the system.In this presentation, I will summarize a few of my contributions on the use and the refinements of mean field approxiamtion to study the performance of distributed systems and algorithms. I will introduce the key concepts behind mean field approximation, by giving some examples of where it can be applied. I will review some of the classical models and try to answer a very natural question: how large should a system be for mean field to apply?
This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.
Strictly Necessary Cookies
Strictly Necessary Cookie should be enabled at all times so that we can save your preferences for cookie settings.
If you disable this cookie, we will not be able to save your preferences. This means that every time you visit this website you will need to enable or disable cookies again.
