Date:  2013.8.1 (Thu)
Time:  14:00 - 18:00
Place: CELC seminar room

Speakers (only guest speakers, some from C01 will give a talk also)

Erik Aurell (KTH, Stockholm)

Title: The inverse Ising problem and its generalizations have
       emerged as a common point of interest 

Abstract:
Learning spatial amino acid contacts from many homologous protein sequences
of statistical physics and machine learning. A major success has been
achieved in determining physical proximity of amino acid residues in
families of homologous proteins by learning an appropriate Potts model
where each variable can take 21 states representing the 20 amino acids
and a gap in an alignment (Lapedes et al 2001 (unpub), Weigt et al 2009,
Marcos et al 2011, Hopf et al 2012 and many follow-up papers).

Improving such contact predictions can be achieved by (at least) (i) starting
from better and/or more appropriate protein families and sequence alignments
(ii) learning the Potts model more accurately and (iii) learning other models
which allow for better prediction of the protein structures. I will review
these points with an emphasis on (ii) and (iii) and discuss why it appears
to make sense to learn such relatively simple models as Potts models to
describe in principle a quite complicated trace left in the molecular
record by evolution.

Our work in this direction is presented in Ekeberg et al Phys. Rev. E 87,
012707 (2013) and our codes are available at plmdca.csc.kth.se.

Pinyang Lu (Microsoft, Shanghai)

Title: Approximate Counting via Correlation Decay

Abstract:
In this talk, I will survey some recent development of approximate counting
algorithms based on correlation decay technique. Unlike the previous major
approximate counting approach based on sampling such as Markov Chain Monte
Carlo (MCMC), correlation decay based approach can give deterministic fully
polynomial-time approximation scheme (FPTAS) for a number of counting
problems.