Statistical and Adversarial Frameworks for Prediction Problems

Machine learning theory has developed a variety of theoretical
frameworks for the design and analysis of learning algorithms. This talk will
introduce two well studied frameworks: statistical and adversarial. In the
former, data is assumed to be generated using a probabilistic mechanism whereas
in the adversarial framework, it could have been generated by a malicious
adversary and is revealed to the learner sequentially. We will see how a
fundamental measure of statistical complexity, called the Rademacher
complexity, has a nice analog in the adversarial framework. The talk will end by
providing pointers to global and local resources for students wishing to enter
the fascinating area of research in machine learning theory.