Adversaries, Extrapolation, and Language
Speaker
Percy Liang
Stanford University
Host
Aleksander Madry
MIT CSAIL
Abstract:
Despite their excellent performance on benchmarks, state-of-the-art machine learning systems generalize poorly out of domain and are vulnerable to adversarial examples. We show this on the SQuAD question answering benchmark and suggest two general directions for improvement. First, working in an unsupervised learning setting can promote the development of models with better inductive biases. Specifically, we show how to learn an end-to-end neural network based on message passing that can solve SAT instances given only instances labeled with whether they have a solution. Second, we show how to use natural language as a means of providing stronger supervision. In one project, we convert natural language explanations to a function that labels unlabeled data, which can be used to train a predictor. In another, users interactively teach high-level concepts using natural language definitions. While being far from a complete solution, we hope that these vignettes are suggestive of broader ideas worth exploring.
Bio:
Percy Liang is an Assistant Professor of Computer Science at Stanford University (B.S. from MIT, 2004; Ph.D. from UC Berkeley, 2011). His research spans machine learning and natural language processing, with the goal of developing trustworthy agents that can communicate effectively with people and improve over time through interaction. Specific topics include question answering, dialogue, program induction, interactive learning, and reliable machine learning. His awards include the IJCAI Computers and Thought Award (2016), an NSF CAREER Award (2016), a Sloan Research Fellowship (2015), and a Microsoft Research Faculty Fellowship (2014).
Despite their excellent performance on benchmarks, state-of-the-art machine learning systems generalize poorly out of domain and are vulnerable to adversarial examples. We show this on the SQuAD question answering benchmark and suggest two general directions for improvement. First, working in an unsupervised learning setting can promote the development of models with better inductive biases. Specifically, we show how to learn an end-to-end neural network based on message passing that can solve SAT instances given only instances labeled with whether they have a solution. Second, we show how to use natural language as a means of providing stronger supervision. In one project, we convert natural language explanations to a function that labels unlabeled data, which can be used to train a predictor. In another, users interactively teach high-level concepts using natural language definitions. While being far from a complete solution, we hope that these vignettes are suggestive of broader ideas worth exploring.
Bio:
Percy Liang is an Assistant Professor of Computer Science at Stanford University (B.S. from MIT, 2004; Ph.D. from UC Berkeley, 2011). His research spans machine learning and natural language processing, with the goal of developing trustworthy agents that can communicate effectively with people and improve over time through interaction. Specific topics include question answering, dialogue, program induction, interactive learning, and reliable machine learning. His awards include the IJCAI Computers and Thought Award (2016), an NSF CAREER Award (2016), a Sloan Research Fellowship (2015), and a Microsoft Research Faculty Fellowship (2014).