Unified Representations for Learning and Reasoning

Yitao Liang

2020-01-01

Abstract:The recent decade has witnessed unprecedented success of deep learning in various high-profile challenges, e.g., the ImageNet recognition task, Atari and the board game Go, human-like text generation, and protein structure prediction, etc. Encouraged by those remarkable empirical achievements, both the research and industry communities cannot wait but further expand AI’s impact in other scenarios. Einstein once famously said “the important thing is not to stop questioning”. However, embracing the recent success, many have started paying less attention to what has exactly been learned and what we can do with what has been learned. Before the practice of “more data, more computation units” runs out of fuel, it is never too late to re-contemplate what is missing from current representations. My research focuses on investigating three such aspects, and proposing promising solutions to address them.