Friday, February 8, 2019 at 12:00pm to 1:30pm
Learning Machines Seminar Series
Additional talk details TBA.
Percy Liang's research is to develop trustworthy agents that can communicate effectively with people and improve over time through interaction. He broadly identifies with the machine learning (ICML, NIPS) and natural language processing (ACL, NAACL, EMNLP) communities.
Agents need to be able to "understand natural language." Much of his work has centered around the task of converting a user's request to simple computer programs that specify the sequence of actions to be taken in response. Recently, he has been exploring agents that learn language interactively (ACL 2017) or can engage in a collaborative dialogue with humans (ACL 2017).
Despite the successes of machine learning, notably deep learning, otherwise high-performing models are still difficult to debug and fail catastrophically in the presence of changing data distributions and adversaries. Given our increasingly reliance on machine learning, it is critical to build tools to help us make machine learning more reliable "in the wild." Recently, he has worked on estimating the accuracyof a predictor on an unknown distribution (NIPS 2016), using influence functions to understand black-box models (ICML 2017), and trying to provide formal guarantees that a learning algorithm is safe from adversaries.
Finally, he is a strong proponent of efficient and reproducible research. He has been developing CodaLab Worksheets in collaboration with Microsoft Research, a new platform that allows researchers to maintain the full provenance of an experiment from raw data to final results. Most of our recent papers have been published on CodaLab as executable papers.