◆  Speaker：Zhouhan Lin

Title: Accelerating LLMs by Exploiting Redundancies in Transformers

Short Bio: Dr. Zhouhan Lin is currently an Associate Professor and Deputy Director at the John Hopcroft Center of Computer Science at Shanghai Jiao Tong University. Before joining SJTU, he was a visiting scientist at Facebook Al Research (FAIR) in Menlo Park, CA. Dr. Lin received his PhD in Computer Science from the Mila lab in the University of Montreal in 2019, where he was supervised by Dr. Yoshua Bengio. His core research interest is to explore and develop machine intelligence capable of acquiring, forming, reasoning, and interacting with abstract concepts from large amounts of data. Dr. Lin proposed the early version of self-attention, which was later integrated into the famous Transformer and Graph Attention Nets. He has accumulated over 9000 citations according to Google Scholar, with two papers having over 2000 citations each, and 10 papers having over 100 citations. Dr. Lin has served as area chair or senior program committee member in several top-tier venues such as EMNLP 2022, AAAI 2021, COLING 2024, and AACL 2022.

Abstract: The redundancies existing in LLMs are an important phenomenon that many methods leverage it to accelerate the computation of LLMs, such as KV cache compression or attention value pruning. In this talk, we'll first present the redundancies in LLM hidden states through the lens of transform-based methods, both in the dimension of sequences and layers. Subsequently, we'll present several related works that leverage these redundancies to accelerate various aspects of LLMs, including the forward process in the encoder, supervised fine-tuning, inference-time acceleration, and KV cache compression.

◆  Speaker：Xiting Wang

Title: Beyond the Obstacles: How to Overcome Difficult Times in Research

Short Bio: Xiting Wang is a tenure-track assistant professor in Renmin University of China. She was previously a principal researcher at Microsoft Research Asia and obtained her Bachelor's degree and Ph.D. from Tsinghua University. Her research interest is explainable and trustworthy AI, and the technologies she developed have been applied in multiple products like Microsoft Bing and Microsoft News. Xiting is an area chair of IJCAI and AAAI, is the archive chair of IEEE VIS, and was awarded Best SPC by AAAI 2021. Two of her papers were selected as the spotlight article by IEEE TVCG (one spotlight each issue).

Abstract: Research can be tough, especially when you feel like you are not making any progress or losing interest in your topic. In this talk, I will share some of the challenges that I faced and discuss practical strategies that have helped me get back on track, such as setting personal goals that truly resonate with me, following my interests, and breaking big tasks into smaller, manageable steps to achieve a state of flow. Let us explore how to stay motivated and engaged, even when the going gets tough.

◆  Speaker：Shunmin Deng

Title: Change and Constancy: A Brief Discussion on Choosing Academic Research Directions

Short Bio: Dr. Shumin Deng is a postdoctoral research fellow at NUS, dedicated to research in NLP, particularly in areas like Reasoning and Planning with LLMs, LLM Agents, Human-AI Interaction, Information Extraction, and Knowledge Graphs. Dr. Deng has published extensively in prestigious international conferences and journals, such as NeurIPS, ICLR, ACL, EMNLP, ICDE, WWW, WSDM, TASLP, and KBS. She has obtained her Ph.D. degree at Zhejiang University, and received the "Outstanding Graduate of Zhejiang Province" award in 2022. She was recognized as the "Outstanding Research Intern of Alibaba Group" in 2020. Dr. Deng serves as an associate editor of TASLP, also contributes as a Session Chair for conferences including EMNLP, AACL, and ICDE, serves as the Publication Chair for CoNLL'23, and participates in program committees for various conferences like NeurIPS, ICLR, ICML, ACL, ARR, COLM, besides being a reviewer for journals like PNAS, TPAMI, ACM Computing Surveys, TBD, TASLP, TALLIP, SCIENCE CHINA Information Sciences, Journal of Software.

Abstract: In the dynamic and ever-evolving field of NLP, choosing an academic research direction can be both exciting and challenging. This talk delves into the art of balancing innovation with enduring principles in the pursuit of meaningful research.
We will explore how to navigate the tension between emerging trends (change) and foundational concepts (constancy), offering strategies for selecting research topics that are both relevant and personally fulfilling.
Drawing upon personal experiences and notable case studies, the discussion will highlight practical approaches to identifying promising areas of study while maintaining a coherent and adaptable research trajectory. Key topics include assessing the longevity and impact of research themes, aligning personal interests with

◆  Speaker：Zhenhua Dong

Title: My industrial research experiences

Short Bio: Zhenhua Dong is a technology expert and project manager of Huawei Noah’s ark lab. He is leading a research team focused on recommender system and causal inference. His team has launched significant improvements of recommender systems for several applications, such as news feeds, App store, instant services and advertising. With more than 40 applied patents and 60 research articles in TKDE, SIGIR, RecSys, KDD, WWW, AAAI, CIKM etc., he is known for research on recommender system, causal inference and counterfactual learning. He is also serving as PC or SPC members of SIGKDD, SIGIR, RecSys, WSDM, CIKM. He translated the book “the singularity is near” into Chinese, named “奇点临近”. He received the BEng degree from Tianjin University in 2006 and the PhD degree from Nankai University in 2012. He was a visiting scholar at GroupLens lab in the University of Minnesota during 2010-2011.

Abstract: The talk will introduce the research style and innovation mechanism of Huawei Noah’s are lab, and personal research experiences in industry. The speaker will compare the industry research and academic research from different perspectives, such application, uncertainty, objectives, promotion latters, collaborations. Finally, some general suggestions will be given to the students.

◆  Speaker：Bin Liang

Title: How to determine whether an "idea" is worthwhile and how to push a work based on this "idea"?

Short Bio: Bin Liang is a Postdoctoral Fellow at The Chinese University of Hong Kong. He received a Ph.D. in Computer Science and Technology from Harbin Institute of Technology, Shenzhen. His research interests lie in natural language processing, dialogue systems, affective computing, and multimodal learning. He has published more than 40 papers in conferences and journals such as ICML, ACL, EMNLP, WWW, TOIS, TAC, etc.

Abstract: In our daily research, there are often many "ideas" that come up, whether they are small, big, imaginative, or brainstorming. However, at this time, we often encounter some questions: Can this "idea" be considered a genuine idea? Is it worth our time and effort to study, investigate, and experiment? Further, if we think it's worth doing, how can we implement this idea and push the related research work? In this talk, based on my own research experience, I will outline how to determine whether an "idea" is worthwhile, and discuss how to push a work based on this "idea".