Yukawa LAB Seminar

December 24, 2015 (co-hosted by IEEE Signal Processing Society Japan Chapter)
  •  Time: 15:00 - 17:30
  •  Place: BLDG 14, Room 217 (DR7), Yagami Campus, Keio University
  •  Presenter: Prof. Konstantinos Slavakis (University at Buffalo)
  •  Title: Signal processing Tools for Big Data Analytics: Foundations, Algorithms, and Recent Trends
  •  Abstract: We live in an era of data deluge. Pervasive sensors collect massive amounts of information on every bit of our lives, churning out large-scale 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. Nevertheless, with such big blessings come big challenges. Workhorse signal processing and statistical learning tools have to be re-examined in today's high-dimensional data regimes. This talk contributes to the on-going cross-disciplinary efforts in data science by covering a list of diverse big data (BD) topics, ranging from the theoretical and statistical foundations of BD analytics, to the very recently developed techniques for clustering high-dimensional, massive amounts of data. The talk comprises two parts: (A) Theoretical and statistical foundations for BD analytics, and (B) clustering large-scale data. Theoretical foundations covered by part A are: High-dimensional statistical signal processing and succinct data representations; compressive sampling, sparsity, and (non-linear) dimensionality reduction; low-rank models, matrix completion, and regularization for under-determined problems. A short course on big tensor data models and factorizations is also provided. Part B focuses on clustering (unsupervised learning) of large-scale data. Starting from the state-of-the-art subspace clustering, this part presents the very recently introduced multi-manifold modeling of low-dimensional structures on Riemannian surfaces. Concepts, algorithms and theoretical guarantees are provided, together with applications on action identification in video sequences, dynamic texture clustering, as well as state identification in brain networks. Further, departing from the popular strategy of randomly sampling sub-populations of data of massive cardinality to efficiently accomplish learning tasks, the second part of the talk introduces a data-driven random sketching and validation methodology for clustering massive amounts of data under a low-computational footprint. Basic principles, algorithms and theoretical guarantees are offered, together with applications on community identification in large-scale social networks, and efficient clustering of data with cardinality/dimensionality exceeding one million points/dimensions.
  •  Biography: Konstantinos (Kostas) Slavakis was born in Thessaloniki, Greece. He received the M.Eng. and Ph.D. degrees in Electrical and Electronic Eng. from Tokyo Institute of Technology, (TokyoTech), Japan, in '99 and '02, respectively. He has been a Japanese Government scholar, a JSPS Postdoc fellow (TokyoTech), and a PostDoc fellow with the University of Athens, Greece. He served as a tenured Assist. Prof. at the Dept. of Informatics and Telecomms., Univ. of Peloponnese, Greece ('07-'12), and as a research Assoc. Prof. at the ECE Dept., University of Minnesota, USA. Currently, he is an Assoc. Prof. at the EE Dept., Univ. at Buffalo, SUNY, USA. His research interests include signal processing, machine learning, and big data analytics. He has served the IEEE Trans. on Signal Processing as both Associate and Senior Area editor. He has also delivered tutorial talks in ICASSP’12, ’14, and '15, as well as EUSIPCO '14 and '15.

December 22, 2015
  •  Time: 9:30 - 10:30
  •  Place: BLDG 25, Room 402, Yagami Campus, Keio University
  •  Presenter: João Pedro Pedroso (Universidade do Porto)
  •  Title: Heuristics for Packing Semifluids
  •  Abstract: Physical properties of materials are seldom studied in the context of packing problems. In this work we study the behavior of semifluids: materials with particular characteristics, that share properties both with solids and with fluids. We describe the importance of some specific semifluids in an industrial context, and propose methods for tackling the problem of packing them, taking into account several practical requirements and physical constraints. Although the focus of this talk is on the computation of practical solutions, it also uncovers interesting mathematical properties of this problem, which differentiate it from other packing problems.

November 2, 2015
  •  Time: 9:00 AM - 10:00 AM
  •  Place: BLDG 25, Room 402, Yagami Campus, Keio University
  •  Presenter: Prof. Badong Chen (Xi'an Jiaotong University)
  •  Title: Robust and Sparsity-Aware Similarity Measures in Kernel Space
  •  Abstract: Similarity measures play significant roles in machine learning and signal processing . In recent years, some new similarity measures were proposed, which are defined as a certain distance in a kernel space. Typical examples include the Information Potential (IP), Cross Information Potential (CIP), Cauchy-Schwartz Divergence, Correntropy, and so on. In particular, the Correntropy as a nonlinear and local similarity measure is directly related to the probability of how similar two random variables are in a neighborhood of the joint space, controlled by the kernel bandwidth, which also has its root in Renyi's entropy (hence the name “Correntropy"). Since Correntropy (especially with a small kernel bandwidth) is insensitive to outliers, it is naturally a robust cost for machine learning. The Correntropy Induced Metric (CIM) as a nice approximation of the l0 norm can also be used as a sparsity penalty in sparse learning. This talk will give an overview of several similarity measures in kernel space, with a particular emphasis on Correntropy. The applications to robust regression, adaptive filtering, deep learning, and causality detection with biomedical signals, will also be discussed.
  •  Biography: Badong Chen received the B.S. and M.S. degrees in control theory and engineering from Chongqing University, in 1997 and 2003, respectively, and the Ph.D. degree in computer science and technology from Tsinghua University in 2008. He was a Post-Doctoral Researcher with Tsinghua University from 2008 to 2010, and a Post-Doctoral Associate at the University of Florida Computational NeuroEngineering Laboratory (CNEL) during the period October, 2010 to September, 2012. He is currently a professor at the Institute of Artificial Intelligence and Robotics (IAIR), Xi’an Jiaotong University. His research interests are in signal processing, information theory, machine learning, and their applications in cognitive science and engineering. He has published 2 books, 3 chapters, and over 100 papers in various journals and conference proceedings. Dr. Chen is an IEEE senior member and an associate editor of IEEE Transactions on Neural Networks and Learning Systems and has been on the editorial boards of Applied Mathematics and Entropy.