Speaker: Prof. Seungjin Choi (Pohang University of Science and Technology, Korea)
Meta-Learning: Few-shot learning and Warm-starting
Deep learning has achieved great success in various tasks, when it is trained with large amount of labeled data. However, it is a challenging task for deep models to learning from a handful of labeled examples. Few-shot learning, the goal of which is to learn from only a few examples in each class label, has been a popular subject in machine learning and computer vision communities. Recently, various attempts, including meta-learning, have been made to tackle the problem of few-shot learning. In this talk, I introduce recent advances in meta-learning, the underlying idea of which is to leverage past experience to learn a prior over tasks, so that it can quickly adapt to a novel task. I begin with proving an overview of meta-learning, explaining a few existing work. Then I introduce my own recent work: (1) MT-nets on gradient-based meta-learning; (2) TAEML, task-specific ensemble of meta-learners; (3) learning to warm-start Bayesian optimization.
|Date||February 8, 2019 (Fri) 09:30 - 11:00|