Course Website: canvas.gatech.edu

Description: At the heart of most machine learning applications today – like advertisement placement, movie recommendation, and node prediction in evolving networks – is an optimization engine trying to provide the best decision with the information observed thus far in time, i.e. the problem of online learning. One must make online, real-time decisions and continuously improve the performance with the arrival of data and feedback from previous decisions. The course aims to provide a foundation for the development of such online methods and for their analysis. At times when these decisions are viewed in the context of the background of the consumers, there is some perceived bias or discrimination. The course will discuss ethical and legal issues that arise due to such unintended consequences of decision-making and mathematical tools to formalize as well as mitigate such effects.

Outline: The course will have four interleaved components:

  1. fundamental theoretical tools for analyzing online methods,
  2. algorithmic techniques for developing computationally efficient methods,
  3. applications to real-world problems, and
  4. discussion around ethical and legal issues that might arise due to a perceived bias or discrimination in the decisions.

We will outline important connections to statistics, probability, optimization, graph theory, game theory and policy. We will also have some guest lectures by practitioners from the industry. Since the topic of the course is fairly novel, we will indicate directions of further research and many open questions.

Credits: 3-0-3

Pre-requisites: 3133 (Optimization) and 2027 (Probability). Prior knowledge of machine learning or regression will be useful but not strictly necessary. We will develop most of the required background in the class, but will expect some mathematical maturity.

Grading: The course will have 3 assignments spread through the semester (10% each), a midterm exam (30%) and a course project (40%). Projects can involve survey and summarization of literature review or attempt to progress the state-of-the-art on a novel research question. Each project team can have at most 2 students. Projects will be graded for the undergraduate level and the graduate level separately.

References: We will post readings specific to various topics and broadly follow the book Introduction to Online Convex Optimization by Elad Hazan. Another helpful reference is the book Prediction, Learning, and Games by Nicolo Cesa-Bianchi and Gabor Lugosi. The course will include selected topics from recent courses on Online Learning Methods in Machine Learning by Sasha Rakhlin and Prediction and Learning: It’s only a Game by Jake Abernethy.