Organizer: CASED, CROSSING and EC SPRIDE
One of the paramount challenges of this century is that of understanding complex, dynamic, large-scale networks. Such high-dimensional networks, including communication, social, financial, and biological networks, cover the planet and dominate modern life. In this talk, we propose novel approaches to inference of information in such networks, using timing that provides rich information for both active and passive learning scenarios.
In the active learning scenario, we briefly discuss active perturbation of timing in a network to facilitate the inference task, with a focus on applications in privacy-aware resource scheduling and network forensics.
In the passive learning model, we present an information-theoretic framework for learning and visualizing causal influences by representing them as a network. In such networks, called directed information graphs, arrows capture direction of influence between entities of interest, represented by nodes. Directed information graphs are a new type of probabilistic graphical model and maybe thought of generalization of Bayesian networks, the most commonly used graphical model in machine learning and statistical inference.
We discuss applications ranging all the way from biological networks to social networks. Specifically, we apply our approach to learning functional relationships in the brain from ensemble of neural spike train recordings and discovering maps of the direction of information flow in the blogosphere by analyzing hyperlinks provided in one media site to others. In the presence of large data, we propose efficient algorithms that identify optimal or near-optimal approximations to the topology of the network of casual influences.