Prof. Xiaoli Li

The proliferation of sensors across industries such as manufacturing, aerospace, and healthcare has created an urgent need for AI-driven time series analytics. This talk delves into recent advancements in AI techniques that enhance predictive maintenance, optimize machine health monitoring, and improve operational efficiency. We explore key challenges and solutions in the following areas: (1) Traditional Feature Engineering and Classification Models – The role of hand-crafted features and conventional machine learning techniques in time series analysis. (2) Self-Supervised Representation Learning – Leveraging contrastive learning to extract meaningful features from unlabeled time series data. (3) Unsupervised Domain Adaptation – Addressing distribution shifts in multivariate sensor data at both local and global levels to improve cross-domain generalization. (4) Model Compression for Edge Deployment – Enhancing AI model efficiency for resource-constrained environments. (5) Finally, we discuss how foundation models for time series analytics can be adapted to downstream applications, enabling broader real-world adoption.

Biography:

Prof. Xiaoli Li is currently a department head (Machine Intellection department, consisting of 100+ AI and data scientists, which is the largest AI and data science group in Singapore) and a principal scientist at the Institute for Infocomm Research, A*STAR, Singapore. He is a technical director of AI for Manufacturing COE (AIMfg) supported by MTI (35m). He also holds adjunct professor position at Nanyang Technological University (He was holding adjunct position at National University of Singapore for 6 years). He is an IEEE Fellow and Fellow of Asia-Pacific Artificial Intelligence Association (AAIA). He has been recognized as Clarivate's Highly Cited Researcher. Xiaoli is also serving as KPMG-I2R joint lab co-director. He has been a member of Information Technology Standards Committee (ITSC) from ESG Singapore and Infocomm Media Development Authority (IMDA) since 2020. Moreover, he serves as a health innovation expert panel member for the Ministry of Health (MOH), expert panel member for Ministry of Education (MOE), as well as an AI advisor for the Smart Nation and Digital Government Office (SNDGO), Prime Minister s Office, highlighting his extensive involvement in key Government and industry initiatives.

Prof. Chong-Yung Chi

Convex optimization (CVXopt), has been extensively applied in sciences and engineering over the last decades. Artificial Intelligence (AI), such as Machine Learning (ML) and Deep Learning (DL), has been pervasive not only in sciences and engineering but also in our daily lives. A specific mathematical model and problem formulation are required for the former while free from any pretraining; meanwhile, optimal or acceptable approximate solutions can always be obtained, together with insightful performance characteristics and unique properties that may be disclosed and used as the guidelines for practical algorithm implementation and development. A big training dataset and tremendous computing costs are frequently required for the latter, thanks to neither a math model nor intricate mathematics; hence, a tractable performance/convergence analysis is essential but still a bottleneck. In this speech, we will address their intriguing fusion (termed CVXopt-aided AI), which demonstrates fantastic learning performance via the following deep image prior (DIP) based AI application instances:

1. DIP-based Unsupervised Hyperspectral Image (HSI) Denoising:  The sparse noise is detected and suppressed by CVXopt, and then the ground truth is recovered using a DIP (a convolutional neural network).  
   2. DIP-based Unsupervised HSI Super-Resolution (HSI-SR): After suppressing the sparse noise by CVXopt, two coupled DIPs (with identical architecture) in parallel are used to capture the utmost essential spectral (spatial) features from a low (high) spatial resolution HSI X (multispectral Y) and guide the generation of abundance tensor G and spectral signature matrix E, respectively, finally yielding the desired tensor HSI-SR Z from G and E.

Biography:

Chong-Yung Chi currently serves as a professor at the Department of Electrical Engineering and the Institute of Communications Engineering at Taiwan Tsing Hua University. He is an IEEE Life Fellow and a Fellow of both the Asia-Pacific Artificial Intelligence Association (AAIA) and the International Artificial Intelligence Industry Alliance (AIIA). He obtained his Ph.D. in Electrical Engineering from the University of Southern California, USA, in 1983. He is currently a Professor of Taiwan Tsing Hua University. He has published more than 240 technical papers (with citations more than 7500 times by Google-Scholar). His current research interests include signal processing for wireless communications, convex analysis and optimization for blind source separation, and related fields. He was an Associate Editor (AE) for four IEEE Journals, including IEEE TRANSACTIONS ON SIGNAL PROCESSING for 9 years, and he has also served as a member of several technical committees of the IEEE Signal Processing Society

Dr. Xun Xu

Industrial visual inspection is challenging due to the rarity of defects, high variability in normal samples, and distribution shifts during testing (e.g., sensor drift, lighting changes). Supervised methods are impractical due to costly annotations, while unsupervised fine-tuning risks performance degradation. We propose data and computationally efficient solutions using low-rank weight decomposition, weakly labeled data, and robust regularization to fine-tune pre-trained models for downstream industrial inspection tasks. To address rare anomalies and improve generalization, we leverage diffusion models to synthesize diverse anomaly samples. The developed approaches balance efficiency, robustness, and adaptability, enabling effective defect detection and image segmentation for metrology in real-world industrial visual inspection settings.

Biography:

Dr. Xun Xu graduated from Queen Mary University of London in 2016. From 2016 to 2019, he conducted postdoctoral research at the National University of Singapore. He is currently a Senior Scientist at the Agency for Science, Technology and Research (ASTAR), Singapore. Dr. Xu's research focuses on data-efficient computer vision, with significant contributions in semi-supervised learning, domain adaptation, active learning, zero-shot learning, and clustering. His work has resulted in over 50 publications in top AI journals and conferences such as TPAMI, IJCV, NeurIPS, ICLR, CVPR, ICCV, and ECCV, as well as two invited book chapters. He has served as an Area Chair for conferences like NeurIPS, ACM MM and BMVC, is an IEEE Senior Member, received the ICIP 2015 Top 10% Paper Award and EPTC 2022 Best Academic Paper Award. Dr. Xu has led two key research and development projects funded by A*STAR and received the A*STAR Career Development Award. He also contributed to multiple industrial vision inspection projects with leading semiconductor companies for defect detection and yield improvement.

Dr. Joey Tianyi Zhou

This talk introduces Agent Trading Arena—a virtual trading platform designed to simulate complex economic systems through zero-sum interactions—as a testbed to evaluate the reasoning capabilities of Large Language Models (LLMs) in dynamic financial scenarios. It highlights three core findings: (1) visualized data formats significantly enhance LLMs’ geometric reasoning and boost overall return by 4%; (2) reflection modules improve long-term strategic decision-making by filtering short-term noise; and (3) a dynamic agent-based system that simulates realistic price fluctuations. This work offers a novel paradigm for evaluating LLMs in high-stakes domains like finance and healthcare, emphasizing the role of visual inputs in cross-disciplinary reasoning.

Biography:

Dr. Joey Tianyi Zhou is Deputy Director at the Frontier Artificial Intelligence Center, Singapore. He received his Ph.D. from Nanyang Technological University and has led numerous national-level research projects. He has published over 150 papers in SCI Q1 journals and CCF A conferences, including AIJ, TNNLS, NeurIPS, and AAAI. He serves as an editor or board member for AIJ, IEEE Transactions, and others, and has been an Area Chair at NeurIPS, ICML, ICLR, AAAI, and IJCAI. He is the IJCAI 2025 Vice Program Chair and has won Best Paper Awards at IJCAI, ECCV, and ACML. He is listed among the world’s top 2% scientists by Stanford University.

Dr. Min Wu

The drug discovery pipeline is a complex, multi-stage process encompassing target identification, hit discovery, lead optimization, and clinical development. While traditional approaches are often time-consuming and costly, artificial intelligence (AI) has emerged as a transformative tool to accelerate these stages. In this talk, we will explore how AI-driven methods can enhance two critical phases of drug discovery: target identification via synthetic lethality (SL) prediction and hit discovery via RNAsmall molecule binding affinity prediction. I will present our recent work on benchmarking machine learning algorithms for SL prediction. In this work, we systematically benchmarked 12 recent machine learning methods for SL prediction, assessing their performance across diverse data splitting scenarios, negative sample ratios, and negative sampling techniques, on both classification and ranking tasks. I will also introduce our deep learning framework called DeepRSMA for predicting RNAsmall molecule binding affinities. Our method demonstrates superior performance over existing tools, enabling the virtual screening of RNA-targeted ligands with high accuracy.

Biography:

Dr. Min Wu is currently a Principal Scientist at Institute for Infocomm Research (I2R), Agency for Science, Technology and Research (A*STAR), Singapore. He received his Ph.D. degree in Computer Science from Nanyang Technological University (NTU), Singapore, in 2011 and B.E. degree in Computer Science from University of Science and Technology of China (USTC) in 2006. He received the best paper awards in EMBS Society 2023, IEEE ICIEA 2022, IEEE SmartCity 2022, InCoB 2016 and DASFAA 2015. He also won the CVPR UG2+ challenge in 2021 and the IJCAI competition on repeated buyers prediction in 2015. He has been serving as an Associate Editor for journals like Neurocomputing, Neural Networks and IEEE Transactions on Cognitive and Developmental Systems, as well as conference area chairs of leading AI and machine learning conferences, such as ICLR, NeurIPS, KDD, etc. His current research interests focus on AI and machine learning for time series data, graph data, and biological and healthcare data.

Dr. Zhenghua Chen

Time Series Data Analytics (TSDA) is essential for a wide range of applications, including machine fault diagnosis and prognosis, and healthcare informatics. However, TSDA faces several critical challenges, such as the difficulty and high cost of annotating time series data, domain shifts caused by environmental changes or device heterogeneity, and the constraints of limited computational resources on edge devices for real-time monitoring. In this talk, I will show some recent works by our team in addressing these challenges. Our overarching goal is to develop generalizable and deployable AI models for TSDA that require less data and are robust across diverse conditions. 

Biography:

Dr. Zhenghua Chen received his Ph.D. from Nanyang Technological University , Singapore, and currently serves as a Senior Scientist, Principal Investigator, and Lab Head at A*STAR. His research focuses on efficient machine learning and its applications in time-series data analysis, aiming to develop scalable and robust models for real-world scenarios. He has published over 100 papers in top journals and conferences, including IEEE TPAMI, TNNLS, T-Cyber, TIE, TII, AAAI, IJCAI, KDD, ICCV, and NeurIPS, and has authored three English-language monographs. His work has been cited more than 10,000 times on Google Scholar. Dr. Chen serves as Associate Editor-in-Chief of Neurocomputing, and holds editorial roles with several IEEE Transactions and other journals. He has been named among the world’s top 2% scientists by Stanford University for five consecutive years, and has received multiple accolades including the IES Prestigious Engineering Achievement Award, IEEE TIM Outstanding Associate Editor Award, and A*STAR Career Development Award. He also led his team to first place in the CVPR 2021 UG2+ Challenge and won Best Paper Awards at IEEE ICIEA and SmartCity conferences. He currently chairs the IEEE Sensors Council Singapore Chapter. 

Prof. Yun Liu

This report addresses the challenges of Few-Shot 3D Point Cloud Semantic Segmentation (FS-PCS), focusing on two key issues in the current setup: foreground leakage and sparse point distribution. These problems significantly hinder the performance evaluation of existing models, highlighting the need for a standardized FS-PCS framework. Moreover, we introduce a novel FS-PCS model that differs from previous methods, which mainly enhance prototypes by optimizing support features. Our approach, based on Correlation Optimization, seeks to improve the correlation between semantic classes and features. Additionally, we explore the role of multimodal information in enhancing FS-PCS. Without increasing the inference cost, we discuss how incorporating multimodal data, such as text and 2D images, during model training can boost the representation of 3D point clouds in few-shot scenarios. In the realm of Generalized Few-Shot 3D Point Cloud Segmentation (GFS-PCS), which is often constrained by the sparse knowledge provided by few-shot samples, we propose a new framework called GFS-VL. This framework capitalizes on the strong generalization capabilities of 3D Vision-Language Models (3D VLMs) for new categories. It combines dense, albeit noisy, pseudo-labels from 3D VLMs with accurate but sparse few-shot samples to leverage the strengths of both. To address the limited diversity of new categories in existing GFS-PCS test benchmarks, we introduce two new benchmarks with more diverse new categories to provide a comprehensive evaluation of the model's generalization ability.

Biography:

Dr. Yun Liu is a professor at Nankai University. He previously served as a senior scientist at the Agency for Science, Technology and Research (A*STAR) in Singapore and conducted postdoctoral research at the Computer Vision Laboratory at ETH Zurich, under the supervision of Professor Luc Van Gool. He obtained his Bachelor's and PhD degrees from Nankai University in 2016 and 2020, respectively, under the guidance of Professor Cheng Mingming. His main research interests include efficient perception and segmentation of images, videos, and point clouds. He has published over 30 papers in top journals and conferences in computer vision and artificial intelligence, with more than 8,000 citations on Google Scholar.