Seminars and Events
Acquiring and Understanding Cross-Task Generalization with Diverse NLP Tasks
Event Details
REMINDER:
Meeting hosts only admit guests that they know to the Zoom meeting. Hence, you’re highly encouraged to use your USC account to sign into Zoom.
If you’re an outside visitor, please inform us at (nlg-seminar-host(at)isi.edu) beforehand so we’ll be aware of your attendance and let you in.
In-person attendance will be permitted for USC/ISI faculty, staff, students only. Open to the public virtually via the zoom link and online.
For more information on the NL Seminar series and upcoming talks, please visit:
https://nlg.isi.edu/nl-seminar/
Humans can learn and perform a new language task more efficiently than machines, when they are provided with either task instructions or only a few examples for that task. We believe such learning efficiency is partly achieved by accumulating past learning experiences, i.e., learning to learn with previously seen tasks. We refer to such capability as cross-task generalization and envision it to be an integral piece towards generalist NLP systems.
In this talk, I will present our recent efforts in acquiring and understanding cross-task generalization with diverse NLP tasks: (1) To build a learning environment for acquiring and evaluating cross-task generalization, we construct NLP Few-shot Gym, a repository of 160 few-shot tasks collected from open-access NLP datasets, converted to a unified text-to-text format, and covering diverse formats, goals and domains. We further introduce CrossFit, a few-shot learning challenge that systematically evaluates an algorithm’s ability to quickly learn new tasks. With these resources, we conduct an empirical analysis with multi-task learning and meta-learning approaches, which provides fresh insights. (2) To better understand how models learn transferable skills to achieve cross-task generalization, we develop task-level mixture-of-expert models that explicitly emulates the behavior of accumulating skills and recomposing them when encountering a new task. Our empirical results suggest that training task-level mixture-of-experts can alleviate negative transfer and achieve better few-shot performance on unseen tasks; further we find that the learned routing decisions and experts partially rediscover human categorization of NLP tasks.
Speaker Bio
Qinyuan Ye is a fourth-year CS Ph.D. student at University of Southern California, advised by Prof. Xiang Ren. Her research interests lie in natural language processing. In particular she is interested in approaches that reduce human annotation efforts, including methods leveraging distant supervision, high-level human supervision (e.g., explanations, instructions), and meta-learning. Prior to USC, she was an undergraduate student at Tsinghua University, majoring in Automation.
The recording for this NL Seminar talk will be posted on our USC/ISI YouTube page within 1-2 business days: https://www.youtube.com/user/USCISI.
Subscribe here to learn more about upcoming seminars: