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

Biography: Tianqing Zhu is currently the Dean of Faculty of Data science at City University of Macau, she is also a Professor and PhD supervisor at the City University of Macau and a recipient of the Young Scientists Fund (Class A) of the National Natural Science Foundation of China (NSFC). Before joining the City University of Macau, Zhu Tianqing was a Lecturer at Deakin University in Australia, an Associate Professor at the University of Technology Sydney, and a Professor at China University of Geosciences (Wuhan). She has led couple of NSFC projects, and undertaken a number of projects funded by the Macao Science and Technology Development Fund (FDCT). Before that, she had served as a member of the Australian Research Council (ARC) College of Experts and led & participated in eight ARC-funded projects. As the corresponding or first author, she has published more than 60 CCF-A rated papers in top venues such as NDSS, S&P, USENIX Security and IEEE Transactions, and over 400 SCI-indexed papers. She serves as a Program Committee member for renowned international conferences including CCS, USENIX and AAAI, and as an Associate Editor for several SCI journals. She has been named among Stanford University's World's Top 2% Scientists and has supervised more than 30 PhD students. She is committed to research in artificial intelligence security, focusing on key issues such as attack and defense for intelligent models, data privacy, and the relationship between security and fairness, with the aim of enhancing the security, privacy protection and output fairness of intelligent models.

Abstract: Traditional machine learning constructs models by organizing knowledge from training data. Machine unlearning acts as its reverse process: it seeks to selectively erase the influence of specified training samples, concepts or memorized content from an already trained AI model, ideally making the model behave as if the target data had never been used during training. Driven by privacy regulations, user data deletion rights, removal of poisoned or biased training content, and safety requirements for large models, machine unlearning avoids the heavy cost of full retraining by developing lightweight forgetting mechanisms. This talk will introduce the core definition and paradigm of machine unlearning, distinguish exact unlearning and approximate unlearning, review representative algorithm families, and discuss key evaluation dimensions: unlearning effectiveness, preservation of model utility on remaining data, computational overhead, and verification challenges. We will also highlight open research difficulties, including knowledge entanglement in deep neural networks, persistent memorization in large language models, and attacks that recover erased information. Finally, we outline real-world application scenarios and future directions in the forgetting capability.

Biography: Mikyoung Lim is a professor in the Department of Mathematical Sciences at KAIST. She received her Ph.D. in mathematics from Seoul National University in 2003. She held postdoctoral and faculty positions at École Polytechnique and Colorado State University, respectively, before joining KAIST. Her research interests include inverse problems, spectral analysis of integral operators, and neural network approximation. She was a plenary speaker at the 2025 Applied Inverse Problems (AIP) Conference.

Abstract: We study neural-network approximation with vector-to-scalar activation functions. Using wavelet frame theory on spaces of homogeneous type, we establish an averaging-kernel condition under which the activation function induces a wavelet system. Combining this with $L^2$-convergence results for wavelet expansions, we obtain convergence estimates for neural approximation, including cases with oscillatory and nonsmooth activation functions.