|Time||Nov 6, 2020, 11:00 a.m.|
|Location||Room HFT-TA 617, Einsteinufer 25, 10587 Berlin|
|Title||On the Relevance-Complexity Region of Scalable Information Bottleneck|
Motivated by data and image classification applications, we study the scalable information bottleneck problem, where the encoder extracts multiple stages of information about the source from the observation. Exploiting successive refinement structure, we assume that each stage outputs further description of the source given the information from all previous stages. For this problem, we first characterize the relevance-complexity region for general memoryless source and observation case. Then, we provide closed-form expressions of the relevance-complexity tradeoff for two examples, i.e. the binary case and the scalar Gaussian case, and further propose a modified Blahut-Arimoto algorithm as a numerical method for the optimization. Finally, some numerical examples are provided to illustrate the merit of scalable information bottleneck and its application to the pattern classification.
Mari Kobayashi received the B.E. degree in electrical engineering from Keio University, Yokohama, Japan, in 1999, and the M.S. degree in mobile radio and the Ph.D. degree from École Nationale Supérieure des Télécommunications, Paris, France, in 2000 and 2005, respectively. From November 2005 to March 2007, she was a postdoctoral researcher at the Centre Tecnològic de Telecomunicacions de Catalunya, Barcelona, Spain. In May 2007, she joined the Telecommunications department at CentraleSupélec, Gif-sur-Yvette, France, where she is now a professor. She is the recipient of the Newcom++ Best Paper Award in 2010, and IEEE Comsoc/IT Joint Society Paper Award in 2011, and ICC Best Paper Award in 2019. She was an Alexander von Humboldt Experienced Research Fellow (September 2017- April 2019) and August-Wihelm Scheer Visiting Professor (August 2019-April 2020) at Technical University of Munich (TUM).