Yu-Chiang Frank Wang
Department of Electrical Engineering, National Taiwan University
Meta-learning, or learning-to-learn, is among the machine learning techniques which aims to design models that can learn new skills or adapt to new environments with a few training examples. In contrast to standard supervised learning which requires training with a large amount of (labeled) data for solving a task of interest, meta-learning improves the learning algorithm itself given the experience of multiple learning episodes. It has been observed that meta-learning tackles a wide range of learning tasks with data and computation limitations, as well as the fundamental issue of generalization. In this talk, we will discuss definitions of meta-learning, followed by its relations to topics like transfer learning, multi-task learning, and hyperparameter optimization. We will take few-shot learning in computer vision as the specific application domain, and point out potential challenges and future research directions.
Keywords：Deep learning, machine learnin, meta-learning, computer vision