Speakers

Tao Qin
MSRA

Yingce Xia
MSRA

Abstract

Many AI tasks are emerged in dual forms, e.g., English-to-French translation vs. French-to-English translation, speech recognition vs. speech synthesis, question answering vs. question generation, and image classification vs. image generation. While structural duality is common in AI, most learning algorithms have not exploited it in learning/inference. Dual learning is a new learning framework that leverages the primal-dual structure of AI tasks to obtain effective feedback or regularization signals to enhance the learning/inference process. Dual learning has been studied in different learning settings and applied to different applications.

In this tutorial, we will give an introduction to dual learning, which is composed by three parts. In the first part, we will introduce dual semi-supervised learning and show how to efficiently leverage labeled and unlabeled data together. We will start from neural machine translation and then move to other applications. In the second part, we introduce dual unsupervised learning, where the training is in a fully unsupervised manner. We introduce unsupervised machine translation and unsupervised image translation. Finally, we introduce dual supervised learning and beyond, which includes dual supervised learning, dual inference and dual adversarial learning. At the end of this tutorial, we propose several future directions of dual learning.

Outline

Metarials