Course Information

About

This course examines recent advances and emerging research directions in deep learning. Topics covered include foundation models and scaling; efficient adaptation and compression; reasoning and inference-time computation; unified multimodal understanding and generation; world models and physical intelligence; retrieval and long-term memory; tool-using agents; mechanistic interpretability and model steering; and evaluation and benchmarking.

The course places particular emphasis on multimodal AI, including models that understand, generate, and reason about images, videos, language, and the physical world. We will study recent research through its central ideas, design choices, experimental evidence, and limitations.

Course format: This is a seminar and project course. During each reading week, we will discuss two papers, with one paper assigned to each lecture. Students will lead presentations, write short paper reviews, and contribute to the discussion. Everyone is expected to read both papers before their respective sessions.

TA-led PS sessions connect the papers to their implementations through code walkthroughs and selected notebook demonstrations. These sessions will focus on the components needed to understand a method and interpret its behavior. PS sessions are ungraded and have no separate programming assignments.

Throughout the semester, students will also pursue a research project individually or in pairs. Dedicated sessions provide opportunities to develop ideas, present progress, receive feedback, and share final results.

There are no exams. Assessment is based on paper presentations, paper reviews, and the course project.

Paper presentations: The structure of the paper presentations follows the format used in Colin Raffel’s COMP790 course at University of North Carolina and Alec Jacobson's CSC2521 course at the University of Toronto. Each lecture students will play a different role that defines how they will read the paper and focus on distinct points of view while preparing for the in-class discussion. See Presentations page for the details. This process is chosen to provide multiple perspectives, a thorough understanding of the concepts, and more importantly way more fun. The aim of this seminar is to bring students to the state of the art in this exciting field.

This class is taught by Aykut Erdem, and intended for graduate students and ambitious undergraduates with a research experience. To get the most out of the class, students should have a strong knowledge in deep learning (such as COMP541 or COMP547) and good programming skills. If you have doubts whether you meet these requirements, please consult the instructor in advance.

Instruction style: During the semester, students are responsible for studying and keeping up with the course material outside of class time. These may involve reading particular book chapters, papers or blogs and watching some video lectures.

Time and Location

Lectures: Tuesday and Thursday at 13:00-14:10 (SNA 104)

PS Hours: Wednesday 17:30-18:40 (CASE Z08)

Communication

The course webpage will be updated regularly throughout the semester with notes, presentations, and important deadlines. All other course related communications will be carried out through KUHub Learn.

Course Requirements and Grading

Grading for COMP550 will be based on

• Course Project (presentations and reports) (45%),
• Paper Presentations (45%),
• Class Participation (10%)

Schedule