EURASIP Statistics, Optimization, and Signal Processing (STATOS)

A series of thematic workshops in memory of Alex B. Gershman

STATOS Thematic Workshop 2016: Machine Learning from Big Data

Friday, September 2, 2016, 1:30pm-7:00pm

Room "MARGIT", 6th floor, Hilton Budapest

Hess András square 1-3, H-1014, Budapest, Hungary

(This is the EUSIPCO 2016 venue; EUSIPCO ends at 12:30pm, Friday Sep. 2).

· STATOS 2016 home

· Call for participation

· Link to registration (http://statos2016.eventbrite.com/)

· EURASIP’s Special Area Team on Signal and Data Analytics for Machine Learning (SiG-DML)

· STATOS 2016 is packed with cutting-edge in-depth 1-hour plenaries by world-class experts in machine learning, big data analytics, and optimization.

· A rare opportunity to listen, learn, and interact with experts in an intense half-day workshop.

· EUSIPCO runs until noon on Friday Sep. 2; STATOS will continue right afterwards in the same venue. Stay overnight for a grand finale!

Keynote Schedule

Fri. Sep. 2, Room "MARGIT", 6th floor, Hilton Budapest

1:30-2:20 Stephane Mallat, Mathematical Mysteries of Deep Neural Networks

2:20-3:10 Merouane Debbah, Big Data for Pro-active 5G Networks

3:10-4:00 Mario Figueiredo, Learning with Strongly Correlated Variables

4:00-4:30 Coffee break

4:30-5:20 Georgios Giannakis, Adaptive Sketching and Validation for Learning from Big Data

5:20-6:10 Romain Couillet, Random Matrices and Machine Learning

6:10-7:00 Alain Rakotomamonjy, Machine Learning and Optimal Transport

Keynote speakers and abstracts

Romain Couillet, Supelec: Random Matrices and Machine Learning

Abstract: Thanks to its efficiently exploiting degrees of freedom in large multi-dimensional problems, random matrix theory has today become a compelling field in modern (multi-antenna multi-user multi-cell) wireless communications and is currently making powerful headway into large dimensional signal processing and statistics. With the advent of the big data paradigm, challenging machine learning questions arise, which we claim random matrix theory can address like no other tool before. In this talk, after a basic introduction and motivation to random matrix theory, we shall discuss our early findings in the theoretical understanding and the resulting practical improvements of kernel spectral clustering and semi-supervised learning for large dimensional data, community detection on large realistic graphs, and shall also briefly discuss neural networks as well as robust statistics applications.

Merouane Debbah, Supelec & Huawei: Big Data for pro-active 5G Networks

Abstract: Operators, vendors and academia have recognized the tremendous potential of large datasets in order to achieve the challenging requirements of future Radio Access Networks (RANs). It is now evident that “Big Data” enables crucial insights in RAN operation, leads to new operation paradigms, and enables a degree of optimization that goes far beyond the reactive and self-centered operation of wireless networks today. New system functions, which heavily rely on context information, user profiling and cross layer communication are now validated and are moving from research to product integration. In this talk, we will discuss the promises and challenges that Big Data will be bring to next generation networks.

Mario Figueiredo, IST: Learning with Strongly Correlated Variables

Abstract: In high-dimensional supervised learning, highly correlated variables are a challenge to variable selection methods and make standard sparsity regularization inadequate. Moreover, in many contexts, it is often important to explicitly select all the relevant variables and identify groups thereof. This talk addresses a regularizer that has been recently proposed for this purpose, reviewing both theoretical guarantees and optimization aspects.

Georgios Giannakis, U. Minnesota: Adaptive Sketching and Validation for Learning from Big Data

Abstract: We live in an era of data deluge. Pervasive sensors collect massive amounts of information on every bit of our lives, churning out enormous streams of raw data in various formats. Mining information from unprecedented volumes of data promises to limit the spread of epidemics and diseases, identify trends in financial markets, learn the dynamics of emergent social-computational systems, and also protect critical infrastructure including the smart grid and the Internet’s backbone network. While Big Data can be definitely perceived as a big blessing, big challenges also arise with large-scale datasets. This talk will put forth novel algorithms and present analysis of their performance in extracting computationally affordable yet informative subsets of massive datasets. Extraction will effected through innovative tools, namely  adaptive censoring, random subset sampling (a.k.a. sketching), and validation. The impact of these tools will be demonstrated in machine learning tasks as fundamental as (non-) linear regression, classification, and clustering of high-dimensional, large-scale, and dynamic datasets. 

Stephane Mallat, ENS: Mathematical Mysteries of Deep Neural Networks

Abstract: Deep convolutional networks have obtained spectacular results for image understanding, audio and medical signal analysis, natural languages... We review their architecture, and analyze their mathematical properties, with many open questions. These architectures seem to linearize important non-linear transformations, while reducing dimensionality with appropriate invariants. They are computed with non-linear contractions, and multiscale linear operators, where wavelets play an important role. Applications are shown for image and audio classification as well as regressions of quantum molecular energies.

Alain Rakotomamonjy, U. Rouen: Machine learning and optimal transport

Abstract: Data are available under different forms and structures and extracting knowledge from them through machine learning tools is of primary importance. Most of these tools are generic in the sense that they do not take advantage of the structure of the data. In this talk, I will discuss the key role that optimal transport can play within the context of machine learning, as optimal transport (OT) theory provides geometric tools to compare probability measures. After introducing the basics of OT, I will show how they can advance the state of the art in some machine learning problems including domain adaptation.

STATOS 2016 Organizing Team

General Chairs

Ignacio Santamaría, Cantabria

N. Sidiropoulos, Minnesota

Technical Chair

Cedric Richard, UNICE

Program Committee

Suleyman Kozat, Bilkent

Paris Smaragdis, UIUC

Nelly Pustelnik, CNRS Lyon

Saikat Chatterjee, KTH

Remi Gribonval, INRIA

K. Slavakis, SUNY Buffalo

Finances

Marius Pesavento, Darmstadt

EURASIP Liaison

Fulvio Gini, U. Pisa

Local arrangements

Abdelhak Zoubir, Darmstadt

Graphic Design

Leda Halatsi, Athens

Co-organized by EURASIP’s Special Area Team on Signal and Data Analytics for Machine Learning (SiG-DML) and the Association for the Promotion of Science and Education, a German nonprofit organization supporting children’s education.