Studies of emerging areas and specialized topics in computer science.
Weather forecasting, predicting stock market trends, and deciding which ads to present on a web page are examples of sequential prediction problems. Online learning is a powerful and popular way of dealing with such problems. An online learning algorithm observes a stream of examples and makes a prediction for each element in the stream. The algorithm receives immediate feedback about each prediction and uses this feedback to improve its accuracy on subsequent predictions. In contrast to statistical machine learning, online learning algorithms donít necessarily make stochastic assumptions about the data they observe, and even handle situations where the data is generated by a malicious adversary. There has been a recent surge of scientific research on online learning algorithms, largely due to their broad applicability to web-scale prediction problems.
This course, Online Learning, will provide a rigorous introduction to state-of-the-art online learning algorithms with an emphasis on algorithm design and theoretical analysis. Topics include: background in convex analysis, regret minimization, online optimization, expert problems, online learning with partial feedback and explore/exploit tradeoffs (a.k.a. bandit problems) and a selection of advanced topics.
Student learning goals
General method of instruction
Class assignments and grading