LogisticsWhere: Zoom Instructor: Yuxiong Wang TA: Rajbir Kataria Updates
OverviewSummary: There has been a recent resurgence of interest in learning to learn, or meta-learning. In the standard machine learning paradigm, a model is trained on a set of examples and is specialized for the single task it is trained for. By contrast, meta-learning is performed on a set of tasks and leverages prior experiences when tackling a new task. This course will cover foundation principles, historical perspective, and recent progress of meta-learning. We will position meta-learning with respect to related areas, such as transfer learning, multi-task learning, and continual learning. The course will also discuss various applications of meta-learning in the fields of computer vision, natural language processing, reinforcement learning, and robotics. Students will be required to present and critique research papers and perform a related research project. By the end of the course, students will be able to understand and implement the state-of-the-art meta-learning algorithms and be ready to conduct research in this direction. Prerequisites: While there are no formal prerequisites, this course assumes familiarity with machine learning (CS 446 or similar) and computer vision (CS 543 or similar). If you have not taken courses covering this material, consult with the instructor. Awards: At the end of the course, we will vote and have prizes for:
People
Academic Integrity Policy
Credits and AcknowledgmentI gratefully thank Abhinav Gupta, Lana Lazebnik, and Bo Li for borrowing much of their course design, thank Chelsea Finn, Sergey Levine, Hugo Larochelle, Oriol Vinyals, and Timothy Hospedales for their slides. |