Speech Title:    Multimodal Pixel-level Scene Understanding in the Wild

Abstract:
In the field of computer vision, achieving a comprehensive understanding of complex visual scenes at the pixel level remains an ongoing challenge. This talk provides a comprehensive overview of state-of-the-art techniques, their applications, and the transformative potential of pixel-level scene understanding across various domains. We will explore foundational concepts and delve into cutting-edge advancements, including fully supervised scene segmentation, few-shot and zero-shot scene segmentation, interactive and language-guided segmentation, and video segmentation in complex environments. Furthermore, the talk will present several new segmentation tasks and their corresponding large-scale datasets that are released in CVPR 2023 and ICCV 2023, such as Generalized Referring Expression Segmentation (GRES) task and its dataset gRefCOCO, complex video object segmentation dataset MOSE, and motion expression video segmentation dataset MeViS. These tasks and datasets not only challenge the boundaries of existing techniques but also foster the future development of pixel-level scene understanding in the wild.

Short Bio:
Henghui Ding is currently a tenure-track Professor at Fudan University. Prior to that, he was a Presidential Postdoctoral Fellow (Principal Investigator) at Nanyang Technological University (NTU) in Singapore, a Postdoctoral Researcher at ETH Zürich in Switzerland, and a Research Scientist at ByteDance AI Lab in Singapore. He received the Ph.D. degree (winner of the Best Thesis Award) from NTU in 2020, and the B.E. degree from Xi'an Jiaotong University in 2016. His research interests broadly include machine learning and computer vision. Specifically, he focuses on scene understanding (e.g., image/video segmentation, detection), multi-modal learning (e.g., vision-language), low-shot learning, and human-computer interaction. He serves/served as an Area Chair of CVPR'(24-25), NeurIPS'24, ICLR'25, ACM MM'24, BMVC'24, and a Senior Program Committee (SPC) member of AAAI'(22-25), IJCAI'(23-25).

Speech Title:    Optimizing and Benchmarking Open-Source LLMs for Enterprise Applications

Abstract:
The rapid evolution of large language models (LLMs) has unlocked transformative potential for enterprise applications, yet deploying these models in real-world settings presents unique challenges. This talk explores cutting-edge strategies for optimizing and benchmarking open-source LLMs tailored for enterprise use. We focus on Retrieval-Augmented Generation (RAG) as a key approach to improving context-aware performance while mitigating issues such as hallucination and output repetition. The presentation introduces the Repetition-Aware Performance (RAP) metric, developed to balance repetition control with model coherence and relevance. Leveraging this framework, we evaluate the impact of tuning parameters like the Repetition Penalty (RPP) across a range of LLMs, from compact models to state-of-the-art architectures. Benchmarking results highlight the role of prompt engineering and hyperparameter tuning techniques in achieving optimal trade-offs between efficiency and accuracy. Additionally, we delve into logical reasoning capabilities using innovative datasets and evaluation metric, showcasing the competitiveness of fine-tuned open-source models compared to proprietary counterparts. This session provides actionable insights into deploying robust, high-performing LLMs in enterprise environments, demonstrating their potential to drive impactful innovations while addressing practical implementation challenges.

Short Bio:
Mr. Donghao Huang serves as the Vice President of Research and Development at Mastercard, where he founded the company's R&D division in Singapore. As the global R&D lead for emerging technologies, Mr. Huang drives the integration of innovative technologies such as blockchain, Web3, quantum computing, privacy-enhancing technologies, generative AI, and machine learning into financial systems. His work enhances secure payment transactions, decision-making, and customer experiences. With a robust background in software development, R&D, and technological innovation across diverse industries, Mr. Huang holds degrees from Tsinghua University, Peking University, and the National University of Singapore. Currently, he is pursuing a Doctor of Engineering degree at Singapore Management University, specializing in Generative AI applications in finance.

Speech Title:    Tutorial on grammatical question-answering from RDF data

Abstract:
In this tutorial Professor Jyrki Nummenmaa introduces the problem of answering complex natural language questions from RDF data. He start with an introduction and a discussion of the LC-QuAD question set and RDF data. He discusses the reason why we use RDF for question answering, and pointing out that currently there are lot of RDF data, for instance, Wikidata is stored as RDF data. Then he introduces the basics of grammatical information and grammatical parsing, followed by how to analyze text grammatically and how we can design applications that utilize the grammatical information. Then, he introduces a grammar-based solution to answer complex natural language questions from the LC-QuAD set using grammatical parsing. Then he analyzes and compares different approaches including both grammatical and LLM based solutions. Finally, he discuss the general benefits of using grammars for question-answering applications.

Short Bio:
Dr. Jyrki Nummenmaa is a professor of computer science at Tampere University, Finland, where he leads the Tampere natural language processing (NLP) group. He obtained a MSc in mathematics in 1989 and a PhD in Computer Science in 1995, both at the University of Tampere. He has 3,5 year work experience of the software industry, and he has worked as a professor at the University of Tampere since 1997.He has over 100 peer-reviewed scientific publications, including both rule-based / grammatical NLP solutions, and approaches based on machine learning. He has done research on natural language processing, data bases, algorithms, and data mining. He was former member of the university's IT management steering committee and has participated in various AI based projects and database-related information systems development projects. He has served various conferences and journals as a reviewer, PC member or program chair.