Recently, Tie-Yan has done advanced research on deep learning and reinforcement learning. In particular, he and his team have proposed a few new machine learning concepts, such as dual learning, learning to teach, and deliberation learning. Dual learning leverages the structure duality of AI tasks to enable effective learning even if there are no sufficient training data. Together with some other innovations including deliberation networks, dual learning has achieved the best performance in many tasks (including human parity in Chinese-to-English news translation, and the first place in 8 tasks of WMT 2019). Learning to teach goes beyond traditional machine learning, and utilizes reinforcement learning technologies to automate the data selection, loss function selection, and hypothesis space selection of machine learning tasks. It enlarges the scope of classical machine learning, and achieved state-of-the-art results in many tasks. These inspiring works were published at NeurIPS, ICML, and ICLR, and attracted a lot of attention from the research community. In addition, Tie-Yan’s team built the world-best Mahjong AI, named Suphx, which achieved 10 DAN on the Tenhou Mahjong platform in mid 2019.