Statistical mechanics of unknown unknowns

Mar 20, 2013, 10:30 AM
132:028 (Nordita)




Prof. Yasser Roudi (NTNU, Kavli Institute for Systems Neuroscience)


Our understanding of complex systems is typically limited by the fact that we only see a fraction of the variables describing these systems: we only see parts of a financial market, we typically only see and analyze subparts of a protein protein interaction network, and we can only record from the activity of few cells in the brain; a potentially large number of variables in these systems are not directly observed and measured. In addition to the existence of unknown (hidden) variables, we face the problem that we cannot even monitor the known (observed) variables for a very long time, that is, we may have under-sampled data. In many cases we may not even know that other variables beyond what we see exist and influence our system, that is we may be dealing with "unknown unknowns". What is the effect of this incomplete data? This question can be posed as a problem of inference in the presence of latent variable, something that people in machine learning are very interested in. In this talk, I will describe some results on how not knowing the relevant variables influences our effort in modeling complex systems. I will describe recent advances in statistical modeling of data in which by using techniques from non-equilibrium statistical physics we can build efficient approaches to take into account the influence of hidden variable. References. B. Dunn, Y. Roudi (2013) in press, arXiv:1301.7275v1 M. Marsili, I. Mastromatteo, Y. Roudi (2013) arxiv 2013 arXiv:1301.3622v2 J. Tyrcha, J. Hertz (2013) arXiv:1301.7274v1

Presentation materials

There are no materials yet.