Review by

Sravan Patchala. Feel free to contact him at sravps7@gmail.com

Review

First things first. Yes, it is another machine learning course. But it has one important difference and that makes it worthy of having a different course for this. In “offline” ML (the one everyone is accustomed to), one assumes all the “train” and “test” data is available for usage all at once. In “online” ML, the assumption is that data samples arrive in sequential rounds. So in each round, the “online” ML algorithm chooses the appropriate “best” action. Designing and analysis of such algorithms is what is covered in this course.

The course covers some interesting algorithms in the beginning. Gradually, the course takes a strong inclination towards performance analysis and proving certain bounds, which might seem slightly uninteresting. Verification of these performance metrics and bounds using empirical data is a typical coding assignment in this course. Finally, the course covers a variety of interesting bandit-problems. The course also has a course project towards the end, which typically involves implementing and improving on a related research paper.

Also, doing this course in parallel with a traditional ML course might be slightly beneficial. It would give a better insight to certain common topics and algorithms. On the whole, the course has a single exam during the midsem period, which is typically an open-book exam. So apart from the midsem, every other graded part of this course is to be done “offline”. So getting a good grade would be easy with periodic efforts.