科学研究

主题:Novelty Detection and Adaptation for Open-world Autonomous AI

主讲人:美国英特尔实验室 Sahisnu Mazumder

主持人:计算机与人工智能学院 杨新

时间:2024年11月07日 11:00-12:00

地点:柳林校区通博楼B514,腾讯会议:196655099

个人简介:Sahisnu Mazumder是美国英特尔实验室AI研究科学家,专注于人机协作、对话和交互系统的研究。加入英特尔之前,他在美国伊利诺伊大学芝加哥分校(UIC)获得计算机科学博士学位,并在印度理工学院(IIT)Roorkee获得硕士学位。他的研究方向包括自然语言处理(NLP)、深度学习、自然语言理解(NLU)、对话与交互系统、终身学习与持续学习、以及开放世界AI与学习等。他于2024年由施普林格出版了一本关于“终身与持续学习对话系统”的书,并在顶级AI、NLP和对话会议上发表了30多篇研究论文。此外,他还在SIGIR-2022、IJCAI-2021和大数据分析(BDA-2014)等会议上讲授了教程,并担任COLING-2022的口头报告环节主席,以及多个顶级AI/ML/NLP会议和期刊的审稿人和程序委员会成员。在攻读博士期间,他还曾在华为研究院(美国)和微软研究院(雷德蒙德)担任研究实习生,从事用户活动与兴趣挖掘及自然语言交互(NLI)系统设计等相关研究。

Biography: Sahisnu Mazumder is an AI Research Scientist at Intel Labs, USA where he works on Human-AI collaboration and dialogue & interactive systems research. Prior to joining Intel, he obtained his PhD in Computer Science at the University of Illinois at Chicago (UIC), USA and master’s from Indian Institute of Technology (IIT) - Roorkee, India. His research interests fall into the area of Natural Language Processing (NLP), Deep Learning, Natural Language Understanding (NLU), Dialogue and Interactive Systems, Lifelong and Continual Learning, Open-World AI/Learning. He has authored a book on “Lifelong and Continual Learning Dialogue Systems” which has been published by Springer Nature in 2024. He has also published 30+ research papers in top-tier AI, NLP and Dialogue conferences; delivered tutorials in SIGIR-2022, IJCAI-2021, Big Data Analytics (BDA-2014) and served as an Oral Session Chair at COLING-2022 and PC Member / Reviewer of premier AI/ML/NLP conferences and journals. During his PhD, he also worked as a Research Intern at Huawei Research USA on projects related to user activity and interest mining and at Microsoft Research - Redmond on Natural Language Interaction (NLI) system design.

主讲内容:随着人工智能(AI)智能体的广泛应用,需要考虑如何让它们更加自主,实现自我驱动的持续学习,并灵活应对新情境。在充满未知和变化的开放、动态的现实世界中,智能体必须具备关键能力,如检测和适应新情况、获取知识并逐步学习。具备这些能力将使AI智能体在长期内变得更加智能、具备自我维持的能力。在本次讲座中,将讨论AI智能体如何在现实环境中检测并持续适应新情况,介绍“在职学习”(on-the-job learning)的新范式,并提出一个名为SOLA的框架,用于构建开放世界中的自主AI智能体,还将探讨新情况检测的研究进展,并分享一些“在职学习的方法”,帮助智能体通过交互式知识获取和部署后学习来适应新的情境。最后,总结当前研究中的挑战与未来的研究机遇。

Abstract: With the increasing proliferation of AI agents, it’s time to think about how to make them autonomous, enabling continuous learning and adaptation to new situations in a self-motivated and self-initiated manner. In the open and dynamic real-world environment, where unknowns and novelties are prevalent, critical capabilities like detecting and adapting to novelties, gathering knowledge, and incrementally learning from them have become fundamental for AI agents to become more knowledgeable, capable, and self-sustaining over time. In this talk, I will focus on how AI agents can detect novelties in real-world and continuously adapt to them. I will introduce the paradigm of on-the-job learning and present a new framework called SOLA for modeling open-world autonomous AI agents. The talk will also address various novelty detection problems and recent advancements in addressing them and cover ideas and techniques related to on-the-job learning that enable agents to adapt to new situations by interactively acquiring knowledge and learn them after deployment. Finally, I will conclude with a discussion of open challenges and opportunities for future research.