Title

Optimization Algorithms for Solving Problems Arising from Large Scale Machine Learning

Abstract

In the contemporary “big data” age, the use of Machine Learning models for analyzing large volumes of data has been instrumental in a lot of current technological development. These models necessitate solving very large scale optimization problems, presenting challenges in terms of developing appropriate solvers. In addition, especially for problems arising from Deep Neural Network architectures, the resulting problems are often nonconvex, and sometimes nonsmooth, giving additional difficulty.

In this talk I present the standard structural elements of this class of problems, and how these structures can be handled with appropriate parallel architectures. I discuss the state of the art in terms of optimization algorithms for this setting and summarize the prognosis for ongoing and future research.

Title

MRAM-based architectures in Spintec

Speaker

François Duhem

Date

25th April 2018

Time/Place

10am,Claude Chappe (room TBC)

Abstract

Non-Volatile Memories (NVMs) have gained traction in the last few years as they are expected to help mitigating the ever growing energy consumption due to leakage in advanced technology nodes. Among emerging NVM technologies, Magnetoresistive Random-Access Memory (MRAM) is considered to be one of the most promising as it reaches performance levels close to those of Static RAM (SRAM) with very high endurance, intrinsic immunity to radiations and good downsize scalability.
Spintec is a laboratory fully dedicated to spintronics research, aiming at bridging the gap between fundamental research and applications with expertise in fundamental physics as well as in device-oriented technologies. In particular, the design team focuses on the development of design tools for the hybrid CMOS/magnetic technology and the evaluation of hybrid non-volatile circuits (FPGA, processors, etc.).
This seminar will discuss ongoing research activities in Spintec with a focus on architecture and IC design.

Passive RADAR measurement using DVB-T receivers and Software Defined Radio processing

Jean-Michel Friedt (Université de Besançon, FEMTO-ST lab)

Abstract
We demonstrate the use of affordable DVB-T receivers used as general purpose software defined radio interfaces for collecting signals from a non-cooperative reference emitter on the one hand, and signals reflected from non-cooperative targets on the other hand, to map the range and velocity in a passive radar application. Issues include frequency and time synchronization of the DVB-T receivers, mitigated by appropriate digital signal processing relying heavily on cross-correlations.

Passive radar uses existing non-cooperative emitters as signal sources for mapping non-cooperative target range and possibly velocity. The attractive features of this strategy is the lack of dedicated broadband source for RADAR application, low cost from the use of existing emitters, and stealth since the operator is undetectable. This measurement technique has become accessible to the amateur with the availability of low cost receivers ideally suited for software defined radio processing. In the framework of passive radar applications, two receivers must be synchronized to record simultaneously the reference channel and the signal reflected by the targets: cross correlation will then finely identify the reference signal delay in the measurement signal and allow for target identification. In the case of moving targets, a brute force approach similar to Doppler compensation in GPS acquisition is applied for the cross correlation to coherently accumulate energy: the range-Doppler maps hint at the distance to the target and its velocity. Most interestingly, in the latter context, clutter (signals reflected from static targets) is separated from the moving target which becomes well visible even in a complex environment. In this presentation, we discuss the details of real time acquisition and signal post-processing for passive radar application, while addressing some of the challenges of diverting DVB-T receivers from their original application. While passive radar has been demonstrated with FM broadcast emitters, analog television emitters, or wifi, we shall here consider the broadband signal provided by digital terrestrial television broadcast signal.

Webpage

http://jmfriedt.free.fr/

Title

Heavy tailed distributions characterisations and examples of applications in channel modeling

Speaker

Associate Professor Nourddine Azzaoui

Date

16th March 2018

Time/Place

14h00 in TD-C.

Abstract

Currently, we are witnessing the proliferation of wireless sensor networks and the superposition of several communicating objects which have an heterogeneous nature. The advent of Internet of Things networks as well as the increasing demand for improved quality and services will increase the complexity of communications and puts a strain on current techniques and models. Indeed, they must firstly adapt to the temporal and spatial evolution and secondly, they must take into account the rare and unpredictable events that can have disastrous consequences for decision-making. This talk provides an overview of the various spectral techniques used in litterature describe a communication channel having an impulsive behavior. This is mainly motivated by the historical success of interactions between probabilities, statistics and the world of communications, information theory and signal processing. The presentation will be divided into two parts: the first is devoted to the synthesis of various developments on alpha-stable variables and processes in a purely mathematical mind. The second part will be devoted to applications in the context of communications. The two sides will combine two fundamentally linked aspects: first, a theoretical approach, necessary for a good formalization of problems and identifying the best solutions. Secondly, the use of these models in real work of channel modelling.

Title

Deep Learning: history, models & challenges, with an application in signal processing and mobile authentification

Abstract

Representation Learning (also known with its more popular title « Deep Learning ») consists in automatically learning layered and hierarchical representation with various layers abstraction from large amounts of data. This presentation will review the history of the field, the main actors and the major scientific challenges. We will first present a brief introduction into common deep models like convolutional neural networks and recurrent networks, before going more into detail of some selected applications in signal processing.

In particular, we present a large-scale study, exploring the capability of temporal deep neural networks in interpreting natural human kinematics and introduce the first method for active biometric authentication with mobile inertial sensors. This work has been done in collaboration with Google, where the first-of-its-kind dataset of human movements has been passively collected by 1500 volunteers using their smartphones daily over several months. We propose an optimized shift-invariant dense convolutional mechanism (DCWRNN) and incorporate the discriminatively-trained dynamic features in a probabilistic generative framework taking into account temporal characteristics. Our results demonstrate, that human kinematics convey important information about user identity and can serve as a valuable component of multi-modal authentication systems.

Biography

Christian WOLF is associate professor (Maitre de Conférences, HDR) at INSA de Lyon and LIRIS UMR 5205, a CNRS laboratory, since 2005. He is interested in computer vision and machine learning, deep learning, especially in the visual analysis of complex scenes in motion: gesture and activity recognition and pose estimation. In his work he puts an emphasis on models of complex interactions, on structured models, graphical models and on deep learning. He received his MSc in computer science from Vienna University of Technology (TU Wien) in 2000, and a PhD in computer science from INSA de Lyon, France, in 2003. In 2012 he obtained the habilitation diploma, also from INSA de Lyon. Since September 2017 Christian is on leave at INRIA, at the chroma work group at the CITI laboratory (“délégation INRIA”).

Abstract
Ultrascale architectures involve large-scale complex systems joining parallel and distributed computing infrastructures joining technology trends (i.e. RISC/ CISIC/ASICS systems) ultrascale software and hybrid models (i.e. cloud/Edge and Fog)  that will be extended in different escenarios. Precisely, this complexity propose special challenges from different points of view: sustainability, scalability, dynamicity, energy-aware, usability, data management, dependability  and more, in a software and hardware relationship. This presentation shows some towards about this ultrscale architectures, observing the different challenges and how it is possible to threat some of them, depending the context a point of view software/hardware and application context.

Bio
Professor Carlos J. Barrios H. is Doctor in Computer Science of the Nice–Côte d’Azur University in France. Researcher in High Performance Computing and Large Scale Architectures, he works in projects associated with hybrid/high performance architectures for science and ultrascale systems involving design, performance evaluation and implementation mechanisms. From 2012, Professor Barrios is the director of the High Performance and Scientific Computing Center of the Universidad Industrial de Santander in Colombia and assistant professor of the same university. At same time, he is the general chair of the Advanced Computing Services for Latin America and Caribbean and he’s involved in different HPC collaboration networks between Europe and Latin America. Contact: cbarrios@uis.edu.co

Title
Clustering and Data Anonymization by Mutual Information

Abstract
In this talk, we first introduce  the Shannon theoretic
multi-clustering problem and investigate its properties, uncovering
connections with many other coding problems in the literature. The figure
of merit for this information-theoretic problem is mutual information, the
mathematical properties of which make the multi-clustering problem amenable
to techniques that could not be used in a general rate-distortion setting.
We start by considering the case of two sources, where we derive
singleletter bounds for the achievable region by connecting our setting to
hypothesis testing and pattern recognition recognition problems in the
information theory literature. We then generalize the problem setup to an
arbitrary number of sources and study a CEO problem with logarithmic loss
distortion and multiple description coding. Drawing from the theory of
submodular functions, we prove a tight inner and outer bound for the
resulting achievable region under a suitable conditional independence
assumption. Furthermore, we present a proof of the well-known two-function
case of a conjecture by Kumar and Courtade (2013), showing that the
dictator functions are essentially the only Boolean  functions maximizing
mutual information.  The key step in our proof is a careful analysis of the
Fourier spectrum of the two Boolean functions. Finally, we study
information-theoretic applications to the problem of statistical  data
anonymization via mutual information and deep learning methods in which the
identity of the data writer must remain private even from the learner.

Joint works with Dr. Georg Pichler (TU Wien, Austria), Prof. Gerald Matz
(TU Wien, Austria),  Clément Feutry (CentraleSupélec, France) and Yoshua
Bengio (Montréal, Canada)

Short biography
Pablo Piantanida received both B.Sc. in Electrical
Engineering and B.Sc. in Mathematics degrees from the University of Buenos
Aires (Argentina) in 2003, and the Ph.D. from Université Paris-Sud (Orsay,
France) in 2007. Since October 2007 he has joined the Laboratoire des
Signaux et Systèmes (L2S), at CentraleSupélec together with CNRS (UMR 8506)
and Université Paris-Sud, as an Associate Professor of Network Information
Theory. He is an IEEE Senior Member, coordinator of the Information Theory
and its Applications group (ITA) at L2S, and  coordinator of the
International Associate Laboratory (LIA) of the CNRS “Information, Learning
and Control” with several institutions in Montréal and General Co-Chair of
the 2019 IEEE International Symposium on Information Theory (ISIT). His
research interests lie broadly in information theory and its interactions
with other fields, including multi-terminal information and Shannon theory,
machine learning, statistical inference, communication mechanisms for
security and privacy, and representation learning.

To be announced

Abstract

As part of the internet of things (IoT), the number of sensor nodes that wish to communicate with each other has exploded and is expected to further increase dramatically. Such an increase of communication devices inherently leads to involved communication and hypothesis testing scenarios, and thus calls for new coding and testing strategies. The talk presents new strategies and corresponding error exponents for different network scenarios, and it proves information-theoretic optimality of the proposed strategies in some cases. Special attention is given to scenarios where information collected at a sensor is desired at multiple decision centres and where communication is multi-hop involving sensor nodes as relays. In these networks, sensors generally compete for network resources, and relay sensors can process received information with sensed information or forward intermediate decisions to other nodes. Depending on the studied error exponents, some of these intermediate decisions require special protection mechanisms when sent over the network. The talk is based on joint work with Sadaf Salehkalaibar, Roy Timo, and Ligong Wang.