Lecture 14: Online Learning

So far we have mostly studied statistical learning theory, which focuses on batch learning under probabilistic assumptions. More specifically, we have mostly assumed that the learning algorithm has access to a training set S of n examples drawn iid from a distribution D, and we would like the algorithm to produce a hypothesis or predictor that has low error (measured in some way) on fresh samples generated from D (i.e. low generalization error). We saw how to prove generalization bounds based on properties like uniform convergence and covering numbers, and we also saw some algorithms like boosting and SVM for the batch setting. Today and for the next several weeks we will study a different learning paradigm called online learning or sequential prediction. This setting makes no probabilistic assumptions and hence does not care much about generalization. Nevertheless we will see connections to statistical learning and to game theory, as well as many new algorithms.