时间：2022年6月7日（周二） 19:30-20:30

Time: 19:30-20:30, Tuesday, June 7, 2022

线上：ZOOM ID 822 1298 6146

Online: ZOOM ID 822 1298 6146

主持人: 西湖大学工学院 张岳博士

Host: Dr. Yue Zhang, Associate Professor, School of Engineering, Westlake University

主讲嘉宾/Speaker：

Prof. Danqi Chen 陈丹琦

Assistant Professor

Department of Computer Science

Princeton University

主讲人简介/Biography:

Danqi Chen is an assistant professor of computer science at Princeton University and co-leads the Princeton NLP Group. Her recent research focuses on training and adapting language models, knowledge representation & reasoning, and developing scalable systems for question answering, information extraction, and conversational agents. Before joining Princeton, Danqi worked as a visiting scientist at Facebook AI Research. She received her Ph.D. from Stanford University (2018) and B.E. from Tsinghua University (2012), both in Computer Science. Danqi is a recipient of a 2022 Sloan Fellowship, a Lawrence Keyes, Jr./Emerson Electric Co. Faculty Advancement Award, faculty awards from Google, Meta, Amazon, Apple, and Salesforce, and paper awards from ACL 2016, EMNLP 2017, and ACL 2022.

讲座摘要/Abstract:

Large pre-trained language models (LLMs) have utterly transformed the field of natural language processing. However, the ever-increasing scale of LLMs has raised a lot of concerns, especially environmental and financial costs, making them out of reach of most academic research labs. In this talk, I will discuss two recent works from my lab and share some thoughts on how we should move forward in the age of large language models. I will first describe a new training approach designed for training language models with memory augmentation. Our approach can enable language models to better leverage long-range contexts or external knowledge at testing time, with very little compute overhead. I will then describe an effective structured pruning method of compressing pre-trained models for downstream use. This approach is computationally cheap and yields a >10x speedup during inference with little accuracy reduction.

讲座联系人/Contact:

Ms. Huaiwei Shi

shihuaiwei@westlake.edu.cn