Date & time: September 19, 15:10 - 17:30
Place: Center for ELC Seminar Room (4F, 404)

Spearers: Dr. T. Kawamoto (Tokyo Inst. of Tech.)
          Dr. M. Vehekapera (Aalto Univ., Finland)

15:10 - opening

15:15 - 16:15
Speaker: Dr. Tatsuro Kawamoto
(Tokyo Institute of Technology, JSPS Research Fellow)
Title: Detectability limit of a spectral clustering method
Abstract: 
A method of spectral clustering is one of the traditional tools for
extracting community structure out of graphs, which makes use of 
eigenvectors which correspond to small eigenvalues of a graph 
Laplacian. In general, estimating the validity of a community 
detection method is an important problem for practical use and has
been gathering significant attention. In order to estimate the 
situation where a method of spectral clustering is valid, we consider
a pair of loosely connected random graphs and solve its second 
smallest eigenvalue/eigenvector problem using the replica method;
it gives the limit of the coupling strength between clusters, 
below which the method correctly detects the clusters. 

16:30 - 17:30 
Speaker: Dr. Mikko Vehekapera (Aalto University, Finland)
Title: Signal recovery and denoising for sparse linear systems
Abstract:
Several problems in engineering can be thought as an instance of 
signal recovery in linear observation systems.  The current 
state-of-the-art scheme for inference given a sparse source is 
so-called approximate message passing (AMP) that can approach the 
Bayesian optimal performance with polynomial complexity under certain 
conditions.  Recently it has been shown, however, that the basic AMP 
algorithm is not able to provide optimal performance if the coupling 
matrix is not constructed from independent entries.  In this talk we 
review some of the recent results regarding iterative signal recovery 
methods for sparse linear systems - including the case when the 
measurement matrix has specific row-orthogonal structure.

END