Title: Non-Asymptotic Bounds for Forward Processes in Denoising Diffusions: Ornstein-Uhlenbeck is Hard to Beat

Time: 11:00am 

Venue: HW312, Haking Wong Building, HKU

Speaker(s): Prof. Aleksandar Mijatović

Remark(s): 

Abstract

"Denoising diffusion probabilistic models (DDPMs) represent a recent advance in generative modelling that has delivered state-of-the-art results across many application domains. Despite their success, a rigorous theoretical understanding of the error within DDPMs, particularly the non-asymptotic bounds required for the comparison of their efficiency, remain scarce. Making minimal assumptions on the initial data distribution, allowing, for example, the manifold hypothesis, this talk presents explicit non-asymptotic bounds on the forward diffusion error in total variation (TV), expressed as a function of the terminal time T.

The talk parametrises multi-modal data distributions in terms of the distance R to their furthest modes and consider forward diffusions with additive and multiplicative noise. The analysis rigorously proves that, under mild assumptions, the canonical choice of the Ornstein–Uhlenbeck (OU) process cannot be significantly improved in terms of reducing the terminal time T as a function of R and error tolerance. Motivated by data distributions arising in generative modelling, the talk also establishes a cut-off like phenomenon (as R →∞) for the convergence to its invariant measure in TV of an OU process, initialized at a multi-modal distribution with maximal mode distance R.

Joint work with M. Bresar."

About the speaker

Prof. Aleksandar Mijatović is a Professor of Probability at the Department of Statistics at the University of Warwick and a Fellow of The Alan Turing Institute in London. Prof. Mijatović was previously a Chair in Probability at the Department of Mathematics of King’s College London, and before that a Reader in Probability at the Mathematics Department of Imperial College London. Prof. Mijatović obtained his Ph.D. in low-dimensional topology at the University of Cambridge, before working in the City of London as a front-office quantitative analyst in Foreign Exchange derivative markets. His research interests are in Probability and its applications, including Stability of Stochastic Systems, Simulation and Monte Carlo Methods, Mathematical Finance, Numerical Stochastics, Data Science & Foundations of Machine Learning. He is also interested in the interactions of Probability with Analysis and Geometry.

Title: Causal Generalist Medical AI

Time: 11:00am 

Venue: HW312, Haking Wong Building, HKU

Speaker(s): De. Hongtu Zhu

Remark(s): 

Abstract

"The rapid evolution of flexible, reusable foundation models is transforming medical science. This lecture introduces Causal Generalist Medical AI (Causal GMAI)—a paradigm that integrates causal inference into generalist AI architectures to enhance
interpretability, robustness, and generalizability in clinical decision-making. Causal GMAI leverages advanced self-supervised, semi-supervised, and supervised learning across highly diverse, multimodal datasets, including medical imaging, electronic health
records (EHR), clinical trials, genomics, knowledge graphs, and clinical narratives, to perform complex downstream tasks with minimal task-specific supervision.

By embedding structural causal reasoning, these models move beyond traditional correlation-based prediction to infer underlying disease mechanisms and counterfactual outcomes, thereby advancing diagnostic precision and personalized medicine. This
lecture will outline the mathematical and technical foundations of Causal GMAI—specifically focusing on causal discovery, counterfactual reasoning, and domain adaptation under covariate shift—alongside its real-world clinical applications. Finally, the lecture will address critical open challenges in regulatory compliance, statistical validation, and multi-center dataset curation required to ensure clinical reliability. Ultimately, this presentation provides a foundational framework for statisticians, data scientists, and AI practitioners to advance the next generation of trustworthy and interpretable medical AI."

About the speaker

"Dr. Hongtu Zhu is the Kenan Distinguished Professor of Biostatistics, Statistics, Radiology, Computer Science and Genetics at the University of North Carolina at Chapel Hill. He is the Fellow of ASA, IMS, AIMBE, and IEEE. He was a DiDi Fellow and Chief Scientist of Statistics at DiDi Chuxing between 2018 and 2020 and held the Endowed Bao-Shan Jing Professorship in Diagnostic Imaging at MD Anderson Cancer Center between 2016 and 2018. He is an internationally recognized expert in statistical
learning, medical image analysis, precision medicine, biostatistics, artificial intelligence, and big data analytics. He received an established investigator award from the Cancer Prevention Research Institute of Texas in 2016, the INFORMS Daniel H. Wagner Prize for Excellence in Operations Research Practice in 2019, the ICSA 2025 Distinguished Achievement Award, the IMS
2027 Medallion award and Lecture, and the COPSS 2025 Snedecor Award. He has published more than 359 papers in top journals, including Nature, Science, Cell, Nature Genetics, Nature Communication, PNAS, AOS, JASA, Biometrika,
and JRSSB, as well as presenting 71+ conference papers at top conferences, including meetings for Neurips, ICLR, ICML, AAAI, IPMI, MICCAI, and KDD. He is the coordinating editor of JASA and the editor of JASA ACS."

Title: Modeling, Understanding, and Interacting with the 3D World

Time: 02:30pm 

Venue: CB328, 3/F, Chow Yei Ching Building, HKU

Speaker(s): Prof. Mengyu Wang

Remark(s): 

Abstract

The rapid rise of large language models has brought AI into people’s daily lives and is reshaping many aspects of society. It is increasingly recognized that AI’s success in the digital domain must be extended to the real 3D world, ultimately enabling robotic AI systems to live and work in physical environments. Achieving this goal requires models that can effectively model, understand, and interact with the 3D world. In this talk, I will present our recent research spanning 3D object generation, dynamic scene understanding, geometric and spatial reasoning, world models, and active vision systems. In particular, I will introduce Stream3D, a scalable framework for streaming and consistent 3D generation from sparse observations; PAGE-4D, a dynamic-aware 4D reconstruction model that jointly estimates geometry and camera motion in dynamic scenes; GeoWorld, a geometry-grounded world modeling framework that improves spatial reasoning and physical consistency in vision-language models; GEM, a geometry-enhanced world model that aligns generative dynamics with structured geometric representations for robotic manipulation; and an active vision system that enables robots to actively perceive the world, improve scene understanding, and increase manipulation success through closed-loop interaction. Together, these works highlight a pathway toward robotic AI systems that can robustly perceive, predict, and act in the real world.

About the speaker

Prof. Mengyu Wang is an Associate Professor with appointments at Harvard Medical School, Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University, Harvard Data Science Initiative, and Broad Institute of MIT and Harvard. Prof. Mengyu Wang has interests spanning generative AI for computer vision, multimodal large language model behaviors and agents, AI for robotics, AI for genomics, and various other AI applications in medicine.

Title: TechTalk – Data Intelligence of Large Language Models

Time: 04:00pm 

Venue: Tam Wing Fan Innovation Wing Two

Speaker(s): Prof. Anatoliy Swishchuk

Remark(s): 

All members of the HKU community and the general public are welcome to join!

Speaker: Professor Reynold C.K. Cheng, Professor and Department Head of AI & Data Science, School of Computing and Data Science, HKU
Moderator: Professor Ben Kao, Professor, Department of AI & Data Science, School of Computing and Data Science, HKU
Date: 18th June 2026 (Thursday)
Time: 4:00 PM
Mode: Mixed (both face-to-face and online). Seats for on-site participants are limited. A confirmation email will be sent to participants who have successfully registered.
Language: English

Database systems, which provide various operations for defining and querying data, enable large-scale AI systems and intelligent applications in various domains. Due to recent advances in large language models (LLMs), automating database operations through code generation has become increasingly attainable. This capability of having data intelligence in LLMs has given rise to a new paradigm—Data-Centric Code Generation (DCCG)—which aims to build systems that can automatically understand, manipulate, and reason over data.

To realize DCCG, I will discuss our team’s effort in building benchmarking systems, including BIRD-SQL, a large-scale Text-to-SQL benchmark on real databases, and SWE-SQL, which gauges the ability that an LLM resolves user SQL issues. These benchmarks, widely used in the industry, reveal hallucination and other issues faced by LLMs. To address these challenges, I will present our work in graph-aware reasoning, SQL correction, and multi-turn tabular data analysis. They aim to evolve LLMs from static code generators into autonomous, trustworthy agents, with data intelligence, that can understand and generate data-driven software systems.

For more details : https://innowings.engg.hku.hk/dataintelligence/

Title: Finite-Sample Likelihood Ratios for Logistic Regression

Time: 03:00pm 

Venue: CB328, 3/F, Chow Yei Ching Building, HKU

Speaker(s): Prof. Nikita Zhivotovskiy

Remark(s): 

Abstract

Likelihood ratio methods are a central tool in statistical inference, but their classical justification is largely asymptotic and local. In regular parametric models, Wilks’ theorem predicts a universal chi-square behavior, suggesting that likelihood ratio confidence sets should behave as if they had a simple dimension-dependent number of degrees of freedom. I will discuss a nonasymptotic theory for the likelihood ratio in logistic regression. The main result shows that, under arbitrary fixed designs, the worst-case finite-sample behavior can be larger than the classical Wilks prediction by a logarithmic factor, and that this loss is unavoidable. The bound holds uniformly over all design matrices and all true parameters, and does not require the maximum likelihood estimator to exist.

About the speaker

Prof. Nikita Zhivotovskiy is an Assistant Professor in the Department of Statistics at the University of California, Berkeley. His research interests lie at the intersection of mathematical statistics, probability, and learning theory. His work focuses on understanding what can be learned from data under minimal assumptions, with an emphasis on finite-sample, non-asymptotic, and distribution-free guarantees for statistical and machine learning problems.