Learning from Limited Labeled Data
Modern machine learning applications have enjoyed a great boost utilizing neural networks models, allowing them to achieve state-of-the-art results on a wide range of tasks. Such models, however, require large amounts of annotated data for training. In many real-world scenarios, such data is of limited availability making it difficult to translate these gains into real-world impact. Collecting large amounts of annotated data is often difficult or even infeasible due to the time and expense of labelling data and the private and personal nature of some of these datasets. This session will discuss several approaches to address the labelled data scarcity. In particular, the session will discuss work on: (1) transfer learning techniques that can transfer knowledge between different domains or languages to reduce the need for annotated data; (2) weakly-supervised learning where distant or heuristic supervision is derived from the data itself or other available metadata; (3) and techniques which learn from user interactions or other reward signals directly with techniques such as reinforcement learning. The discussion will be grounded on real-world applications where we aspire to bring AI experiences quickly and efficiently to everyone in more tasks, markets, languages, and domains.