Invited Talks
Invited Talks
From anticipating failures in real-world AI systems to explaining multi-omics models, these talks represent the frontline challenges of making deep learning reliable and interpretable in biomedical research.
Distinguished Professor & Yushan Young Scholar
A. K. Sangaiah, Ph.D.
Generative AI-Driven Optimization Intelligence for Revolutionizing Employee Workforce Efficiency
Professor, School of CS & Engineering
Hongmin Cai, Ph.D.
Unsupervised Clustering Methods and Applications for Genes, Cells and Tissues
Associate Professor & Deputy Head of School Research
Leo Zhang, Ph.D.
When AI Goes Wrong: Building Trustworthy Systems Before They Fail
Distinguished Professor, AI Research Institute
Ming-Wen Shao, Ph.D.
Alternative Attacks and Restorative Attacks on Data-free Deep Neural Networks
Associate Professor, School of CS & Engineering
Yuheng Jia, Ph.D.
Reliable Weakly Supervised Learning
Professor, Department of CS & Engineering
Walid Gomaa, Ph.D.
Trustworthy Multimodal Deep Learning for Healthcare and Human Activity Understanding
Associate Scientist, Department of Biostatistics
Weiqiang Zhou, Ph.D.
Deep Learning of Single-Cell Multi-Omics Data
Professor, Faculty of Engineering
Le Ou-Yang, Ph.D.
Interpretable Deep Learning for Omics Data Analysis
Associate Professor, Department of CS
Rizwan Qureshi, Ph.D.
Responsible Large Multi-Modal Models for Biomedical Data Analysis
Trustworthy Multimodal Deep Learning for Healthcare and Human Activity Understanding
Abstract
Deep learning has shown strong potential in biomedical informatics and human-centered healthcare, including activity recognition, rehabilitation assessment, gait analysis, and remote monitoring. This talk examines trustworthy AI for multimodal human activity understanding using visual, thermal, depth, skeletal, and inertial sensing data. Drawing on applications in exercise assessment, wearable-sensor analysis, and human performance evaluation, it highlights practical challenges such as sensor variability, inter-subject differences, noisy observations, dataset bias, and limited model explainability. The talk emphasizes integrating accuracy, transparency, robustness, and human-in-the-loop validation to move healthcare AI from black-box prediction toward auditable and clinically useful decision-support tools.
Biography
Walid Gomaa is a Professor of Computer Science and Engineering at Egypt-Japan University of Science and Technology, holding a Ph.D. from the University of Maryland College Park. His research encompasses human activity recognition, Arabic natural language processing, video content analytics, time-series analysis, and theoretical computer science. His work includes analyzing inertial motion data for healthcare applications such as fall detection, rehabilitation, and remote patient monitoring. He has published extensively in top-tier venues and served as editor, reviewer, and organizer for international journals and conferences.
Generative AI-Driven Optimization Intelligence for Revolutionizing Employee Workforce Efficiency
Abstract
Generative AI has emerged as ground-breaking technology addressing challenges across various domains including Human Resource Management. This talk covers advanced GAN incorporating self-attention mechanisms and metaheuristic optimization algorithms. MetaSEGAN excels in generating high-quality synthetic data, balancing HR datasets, and improving attrition model predictions. Additionally, multilingual Large Language Models demonstrate immense potential in identifying stress indicators by analyzing employee communication data, promoting inclusivity and scalability for global organizations. Future directions involve integrating MetaSEGAN and multilingual LLMs into an intelligent HR chatbot system.
Biography
Arun Kumar Sangaiah received his Ph.D. from VIT University in 2014 and is currently a Distinguished Professor and Yushan Young Scholar at the National Yunlin University of Science and Technology, Taiwan. He has published over 300 research articles, edited 11 books, and led major projects funded by Taiwan's NSTC and international collaborations. He was recognized as a Clarivate Highly Cited Researcher (2021-2023) and Top 2% Scientist by the Stanford Report. His Google Scholar profile records over 28,000 citations with an h-index of 92.
Alternative Attacks and Restorative Attacks on Data-free Deep Neural Networks
Abstract
Deep neural networks are vulnerable to attacks. Black-box attacks utilize only the outputs of the target network, which is more challenging than white-box attacks. The main solution is training alternative networks similar to the target, then using adversarial samples to attack the black-box network. However, we often don't have access to original training data. This talk explores alternative attack methods and restorative attack methods on black-box models under no-data and few-sample conditions, as well as attack methods on 3D Gaussian Splatting models.
Biography
Ming-Wen Shao is a Distinguished Professor at the Artificial Intelligence Research Institute, Shenzhen University of Advanced Technology. He obtained his Ph.D. from Xi'an Jiaotong University in 2005 and completed postdoctoral research in Control Science and Engineering at Tsinghua University. He is an active member of the Chinese Association for Artificial Intelligence and the Chinese Computer Federation, serving on multiple technical committees. He has led five national and six provincial-level research projects and serves as associate editor for Knowledge-Based Systems.
Unsupervised Clustering Methods and Applications for Genes, Cells and Tissues
Abstract
Complex disease emergence is an evolution process from micro to macro. Unsupervised clustering methods can analyze evolution patterns from three scales: genes, cells, and tissues. Traditional methods face challenges with single-omics sequencing data separation, multi-omics network misalignment, and multi-modal imaging registration. We developed single-omics high-dimensional clustering, multi-omics network clustering, and multi-modal collaborative clustering methods. At the gene level, single-omics sequences identify differential genes; at cellular level, multi-omics networks quantify spatial regulation and mine screening targets; at tissue level, multi-modal images identify common lesions—providing accurate whole-field analysis for tumor diseases.
Biography
Hongmin Cai is a Professor at the College of Future Technologies, South China University of Technology, and a visiting professor at Kyoto University. He obtained his Ph.D. in Applied Mathematics from Hong Kong University in 2015 and serves as associate editor for IEEE Transactions on Emerging Topics in Computational Intelligence and Fundamental Research. His publications appear in IEEE TPAMI, Nature Communications, and IEEE Transactions on Medical Imaging. He has led over 10 projects including the National Science Fund for Distinguished Young Scholars, holds 20 authorized patents, and received the first prize from the Guangdong AI Industry Association.
When AI Goes Wrong: Building Trustworthy Systems Before They Fail
Abstract
As AI systems move from controlled benchmarks into real-world decision-making, trust can no longer be reduced to test-set accuracy. Modern AI systems may fail under distribution shift, be manipulated through adversarial inputs or poisoned data, leak sensitive information, or behave unpredictably when connected to tools and external context. These risks are particularly important in biomedical informatics, where data are sensitive, labels are noisy, deployment environments vary across institutions, and model outputs may influence diagnosis, triage, discovery, or clinical workflows. This talk examines how AI systems can go wrong both before deployment and after they are embedded into real systems. Drawing on examples from adversarial machine learning, backdoor and poisoning attacks, privacy leakage, machine unlearning, jailbreaking, and agentic AI, it explains why reliability and interpretability should be treated as system-level requirements. These failure modes are then connected to biomedical deep learning, including robustness under site and population shift, interpretable model behaviour, privacy-aware learning, and trustworthy biomedical foundation models and agents. The central message is that trustworthy AI should be built with domain experts before systems fail in the wild: failures need to be anticipated, measured, explained, and governed, not merely observed after deployment.
Biography
Leo Zhang is an Associate Professor and the Deputy Head of School Research in the School of Information and Communication Technology at Griffith University, Australia. He received his Ph.D. from the Department of Electrical Engineering at City University of Hong Kong in 2016, where he subsequently worked as a Research Fellow in the Department of Computer Science before beginning his academic career in Australia. Prior to joining Griffith, he was a faculty member at Deakin University from 2018 to 2023. His research spans cybersecurity, trustworthy AI, and applied cryptography, with a recent focus on adversarial robustness, poisoning and backdoor attacks, privacy leakage, machine unlearning, and agentic system security. He has published over 100 refereed papers in leading journals and conferences, including IEEE TIFS, IEEE TDSC, IEEE S&P, USENIX Security, NDSS, ICLR, ICML, NeurIPS, CVPR, ICCV, and KDD. He chairs the Secure Learning Task Force in the Neural Networks Technical Committee of the IEEE Computational Intelligence Society and serves as an Associate Editor for IEEE TDSC, IEEE TMM, and IEEE TCYB.
Reliable Weakly Supervised Learning
Abstract
Weakly supervised learning leverages massive unlabeled data with limited labeled data to build machine learning models cost-effectively. However, weak supervision may involve noise, making it crucial to improve model reliability. This talk introduces two critical topics: intrinsic data representation and weak supervision enhancement. We present works on theories and application techniques in these directions, focusing on methods for learning from noisy labels and partial annotations while ensuring model reliability and robustness.
Biography
Yuheng Jia is an Associate Professor at the School of Computer Science and Engineering, Southeast University. He received his Ph.D. from City University of Hong Kong in 2019, where he was advised by Professor Sam Kwong. His research interests include semi/weakly/un-supervised learning, tensor/matrix optimization, and reliable and trustworthy machine learning. He has published more than 70 peer-reviewed papers in top-tier journals and conferences.
Deep Learning of Single-Cell Multi-Omics Data
Abstract
Recent advancements in single-cell sequencing enable measurement of multiple modalities in individual cells, offering insights into the transcriptome and regulome in unprecedented resolution. However, effectively using ultra-high-dimensional multi-omics data to understand gene regulation remains challenging. Inspired by adapting large language models into genomics, we developed scMBERT, a BERT framework-based pretrained deep learning model using single-cell multi-omics data. scMBERT increases model flexibility and performance in downstream tasks like cell type annotation and batch-effect correction, demonstrating the potential of leveraging multi-omics data to improve single-cell genomic analyses.
Biography
Weiqiang Zhou is an Associate Scientist in the Department of Biostatistics at Johns Hopkins Bloomberg School of Public Health, holding a Ph.D. from City University of Hong Kong (2013). His research focuses on developing statistical and computational methods for integrative analysis of large-scale genomic data, with emphasis on gene regulation in cancer and complex diseases. Notable contributions include BIRD for predicting chromatin accessibility and SCRAT/SCATE for single-cell genomic data analysis. He has authored over 50 publications with more than 1,000 citations and received the Faculty Innovation Fund award.
Interpretable Deep Learning for Omics Data Analysis
Abstract
Explainable artificial intelligence (xAI) has achieved remarkable progress in bioinformatics, particularly in omics data analysis. By integrating complex data from genomics, proteomics, and metabolomics, these models have significantly advanced disease diagnosis, treatment planning, and personalized medicine. Utilizing techniques such as feature correlation analysis and visual explanations, xAI has enhanced analytical capabilities and improved model interpretability. Despite these advancements, xAI still faces challenges in clinical practice, including data privacy protection, model generalization, and standardization of interpretation methods. Future research addressing these challenges will further facilitate xAI application in omics data analysis, providing more reliable and transparent AI tools for biomedical research and clinical practice.
Biography
Le Ou-Yang is a Professor at the Faculty of Engineering, Shenzhen MSU-BIT University, holding a Ph.D. from Sun Yat-sen University (2015). He is a Senior Member of the China Computer Federation and has received the Guangdong Outstanding Young Scholar and Shenzhen Excellent Young Scholar awards. He has published over 70 papers in journals including Nature Biotechnology, IEEE TCYB, and Bioinformatics, and serves as Associate Editor for IEEE Transactions on Computational Biology and Bioinformatics. He is a program committee member for major conferences including AAAI, IJCAI, ICML, NeurIPS, and ICLR.
Responsible Large Multi-Modal Models for Biomedical Data Analysis
Abstract
The integration of large multi-modal AI models in biomedical informatics has paved the way for breakthroughs in disease diagnosis, treatment planning, and healthcare insights. This talk explores responsible multi-modal models designed to handle diverse biomedical data—imaging, clinical records, and genomics—while ensuring interpretability and reliability. We discuss strategies for enhancing model transparency, addressing ethical concerns, and a novel scoring function for assessing bias, safety, and fairness. Examples include medical image classification, patient-oriented disease profiling, and cancer-drug response prediction.
Biography
Rizwan Qureshi is a Senior IEEE Member and Associate Professor at Salim Habib University, Pakistan, holding a Ph.D. from City University of Hong Kong. He has held research positions at MD Anderson Cancer Center, Qatar Foundation, the Chinese Academy of Sciences, and the University of Central Florida. His research spans AI, bioinformatics, and computer vision, with a focus on responsible AI in biomedical applications. He serves as Associate Editor at BMC Bioinformatics, with contributions cited over 1,500 times emphasizing interpretability, fairness, and ethical AI.