Frontier of Artificial Network
A Series of Invited Talks @ FAN Group, CityU

Biography: Lisha Chen’s research develops mathematical and algorithmic foundations for large-scale signal processing and machine learning, with a focus on multi-objective and multi-level optimization and learning theory. Her work addresses structured learning problems involving competing objectives, hierarchical decision-making, and constraints arising from real-world scientific and engineering systems. Bridging theory and practice, her research has contributed to IBM products and patents and provides foundational tools for building machine learning systems that can reason over multiple objectives, and adapt to new deployment conditions. Dr. Chen has organized and delivered tutorials on multi-objective learning at ICASSP and AAAI, and served as a reviewer or program committee member for leading conferences and journals in machine learning, signal processing, and optimization. Her achievements have been recognized through the IEEE Signal Processing Society Scholarship, and the IBM AIRC Fellowship.

Abstract: Many real-world learning systems must balance multiple competing objectives while respecting user-specified preferences. This talk presents recent work on preference-guided multi-objective learning, with a focus on how to efficiently navigate and explore trade-offs along the Pareto front. I will first discuss a framework for navigating the Pareto set toward user-preferred solutions, allowing preferences to directly guide optimization. I will then introduce SURF, a complementary method that steers scalarization weights to uniformly traverse the Pareto front, improving coverage of diverse trade-off solutions. The methods have been applied to multi-lingual speech recognition, multi-objective reinforcement learning, and multi-objective LLM alignment. Together, these works provide principled and computationally tractable tools for incorporating user preferences into modern multi-objective learning systems.

Biography: Xiaolin Hu is an Associate Professor in the Department of Computer Science at Tsinghua University. He received his Ph.D. degree in Automation and Computer-Aided Engineering from The Chinese University of Hong Kong in 2007, and subsequently conducted postdoctoral research in the Department of Computer Science at Tsinghua University. He has been a faculty member at Tsinghua since 2009. His research areas include artificial neural networks and computational neuroscience, with main interests in developing brain-inspired computational models and uncovering the brain's mechanisms for processing visual and auditory information. He has published over 100 papers in international journals and conferences, with more than 29,000 Google Scholar citations. He is currently an associate editor for IEEE Transactions on PAMI, and has previously served as an associate editor for IEEE Transactions on Image Processing and IEEE Transactions on Neural Networks and Learning Systems.

Abstract: Normal individuals are able to engage in smooth conversations with others in noisy environments, such as cocktail parties—a phenomenon known as the "cocktail party effect." Current speech recognition systems still perform far worse than humans in such environments. To improve the robustness of speech recognition systems, a common strategy is to first perform speech separation to isolate the speech of the target speaker before recognition. By investigating the brain mechanisms underlying the cocktail party effect, our research group has introduced macroscale brain structures into artificial neural networks, including brain regions such as the visual cortex, auditory cortex, and thalamus, as well as the bottom-up, top-down, and lateral neuronal projections among them. This has led to several novel models that achieve promising results in both audio-only and visually assisted speech separation tasks.

Biography: Jinwei Zhang received his B.S. in Optics and Optical Engineering (2016) and a dual B.S. in Computational Mathematics (2017) from Sun Yat-sen University, China, and his Ph.D. in Biomedical Engineering from Cornell University (2023). He is currently a Postdoctoral Fellow in Electrical and Computer Engineering at Johns Hopkins University and will join the University of Massachusetts Amherst as a tenure-track Assistant Professor in Biomedical Engineering in Fall 2026. His research focuses on accelerated quantitative magnetic resonance imaging, quantitative susceptibility mapping, and lesion-centric imaging analysis in multiple sclerosis and related neurological disorders. He is a recipient of a U.S. Department of Defense Early Investigator Research Award and a Johns Hopkins Data Science and AI Demonstration Award, supporting his efforts to develop and translate lesion-centric imaging biomarkers for neurological diseases.

Abstract: Quantitative magnetic resonance imaging (qMRI) enables noninvasive characterization of tissue microstructure for studying neurodegeneration, but its clinical utility is limited by long scan times and the lack of methods for identifying and tracking individual white matter lesions (WMLs) over time—a hallmark of many neurological disorders. This work develops accelerated qMRI acquisition and reconstruction methods, robust quantitative susceptibility mapping (QSM), and lesion-wise segmentation and tracking algorithms to enable fine-grained analysis of lesion evolution. By integrating pulse sequence programming, physics-informed modeling, and domain-generalizable deep learning, this framework aims to translate lesion-level imaging biomarkers into scalable tools for neurodegenerative disease research and clinical studies.

Biography: Adeel Razi is a Professor of Computational Neuroscience at the School of Psychological Sciences, Monash University Australia, and affiliated with the Turner Institute for Brain and Mental Health and Monash Data Futures Institute. He leads a highly cross-disciplinary laboratory performing research combining engineering, physics, and machine-learning approaches to answer questions that are motivated by and grounded in neurobiology. He develops statistical methods for analysing neuroimaging time series data with a special focus on state-space modelling, Bayesian statistics, and dynamical systems theory. He has made fundamental contributions to the development and application of Dynamic Causal Modelling, especially for resting-state functional MRI, which is a framework for in-vivo investigating of the function of the human brain. His research has implications for building new neuroscience-inspired artificial intelligence systems, treatment of brain diseases and development of new neuro-technologies. His work has been published in journals such as Nature, Nature Human Behaviour, Nature Mental Health, Nature Reviews Neuroscience, Neuron, Nature Communications, and Proceedings of the National Academy of Sciences, and has been featured in The Guardians, BBC, CNN, The Age, The Australian etc. He joined Monash after finishing his postdoctoral studies (2012-2018) at the Wellcome Centre for Human Neuroimaging, UCL, UK. He received the B.E. degree in Electrical Engineering, with a University Medal, from the N.E.D. University of Engineering & Technology in Pakistan, the M.Sc. degree in Communications Engineering from the University of Technology Aachen (RWTH), Germany, and the Ph.D. degree in Electrical Engineering from the University of New South Wales, Australia in 2012.

Abstract: Synthetic biological intelligence (SBI) provides a new way to study learning and adaptive intelligence by embedding living neuronal networks in closed-loop interactive environments. Recent studies show that neuronal cultures can self-organize, learn task-relevant behaviours, and adapt to sparse sensory feedback in real time. These systems offer a valuable contrast to conventional reinforcement learning, especially in terms of sample efficiency and adaptive flexibility. In this talk, I will introduce SBI platforms such as DishBrain as experimental testbeds for understanding the computational principles of biological learning and for developing brain-inspired artificial intelligence. I will discuss recent work that combines active inference, dynamical systems modelling, and experiment-informed generative models to characterize how neuronal cultures interact with virtual environments. These studies provide insights into how adaptive behaviour can emerge from sparse feedback, prediction, and self-organization in living systems. More broadly, SBI points to new directions for AI beyond scale-driven deep learning, including efficient learning, embodied intelligence, continual adaptation, and neuromorphic computation.