AlbaNova Colloquium

Cristopher Moore, "The physics of inference, phase transitions, and the detection of communities in networks"

Europe/Stockholm
Description

There is a deep analogy between Bayesian inference and statistical physics. When we fit a model to noisy data, we can think about the “energy landscape" of possible models, and look for phase transitions where the ground truth suddenly gets lost in this landscape — either because of thermal equilibrium (where “heat” is noise) or because of dynamics (e.g. a metastable state trapped behind a free energy barrier). I’ll use this framework to describe phase transitions in community detection in networks, where communities suddenly become hard or impossible to find. If time permits, I’ll discuss related spectral algorithms, and give a hint of similar phase transitions in other inference problems.