I am looking for Ph.D. students and postdocs. If you are interested in working with me, please, read the following information and send me an email. Your email should briefly convince me that you meet the position requirements and checked what is my research field. General information about the research in my group Nowadays the rate and volume of information flow are sharply increasing and forcing numerous domains to switch from offline to online information processing. Online learning is a subfield of machine learning providing the intelligence behind many online data processing systems. It has already revolutionized personalization and advertising on the Internet and rapidly penetrates many other domains. Online learning proceeds through interactions between the algorithm and environment. Online learning problems are characterized by three major paratemters: (1) the amount of feedback that the algorithm receives on every interaction round (full information; limited (bandit) feedback; partial monitoring; etc.); (2) the environmental resistance to the algorithm (i.i.d. (stochastic); adversarial; etc.); and (3) the structural complexity of the problem (stateless; with side information; Markov decision processes (MDPs); combinatorial structures; etc.). Together these tree parameters define the space of online learning problems (see the figure on the right). Most "classical" online learning algorithms address isolated points in the space of online learning problems. It means that they operate under narrow sets of assumptions that precisely define the problem setting. For example, "prediction with expert advice", "adversarial multiarmed bandits", and "stochastic multiarmed bandits" correspond to the three red crosses on the figure (in clockwise direction). However, these narrow assumptions are rarely satisfied in practice. If the assumptions are stronger than reality, the algorithms are unable to exploit the simplicity of a problem and perform suboptimally. If they are weaker than reality the algorithms are prone to failure. In other words, the "classical" algorithms can exploit simplicity only if it was assumed and accurately described a priori. The research of my group focuses on the new generation of online learning algorithms that address ranges of problems in the space of online learning problems (depicted by dashed lines in the figure). Examples include algorithms that interpolate between full information and bandit feedback (Seldin et al., 2014); algorithms that interpolate between i.i.d. and adversarial environments (Seldin & Slivkins, 2014); and algorithms that interpolate between stateless bandits and bandits with side information (Seldin et al., 2011). The new algorithms exploit the simplicity along the corresponding axes without assuming it a priori. They open a new era in online learning research, where the researchers progress from studying isolated points in the space of online learning problems to studying ranges of problems. The new group members will work on development of algorithms for new ranges of problems, their analysis in terms of upper and lower bounds, and application to reallife problems. Information for PhD students You need to have the following qualifications in order to join the group:
Furthermore, the following qualifications are an advantage:
Information for PostDocs If you are interested in joining the group you need to have the following qualifications:
Furthermore, the following qualifications are an advantage:
